microsoft futuredecoded optics for the cloud · 2/11/2018 · passive component (no transistors or...
TRANSCRIPT
Microsoft Future Decoded
Optics for the Cloud
02112018
Welcome
Optics for the Cloud ndash
How Microsoft Research are re-thinking
future Data Centre technology
Azure Infrastructure Region
le 60 km
Data centres deployed in 50+ regions 10000s of kilometres of fibre
10-20 data centres region
~100000s servers data centre3
How do we scale cloud capacity
hellip
Faster servers
CPU DisksMem
GPUFPGA
1001101010
2013
1001101010
2020
44 ZB 44 ZB
0100010101
1101101010
1011000110
0100010101
1001101010
0010011011
1011101010
0001001110
0110001011
Source IDC 2014
Data doubling every two years
Need for low latency at scaleML and AI driving this further
Slide credit Mark Russinovich
How big is one Zettabytebull 1 ZB = 1021 bytes = 1 Billion TB
bull 10 TB HDD
bull Need 100 Million of those
bull Thatrsquos a 2500 km high tower
bull Data also needs to be transmitted across the network
bull 100 Gbps data centre network link
bull Thatrsquos 2500 years for transmission
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Hyperscale cloud hardware design in
collaboration with the Open
Compute Project (OCP)
Pushing compute
growth
Project Olympus Rack SystemProject Olympus
Custom storage racks optimised for
cold storage
Pushing storage growth
Disk dPower
for d
Cooling
for d
Project Pelican
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Welcome
Optics for the Cloud ndash
How Microsoft Research are re-thinking
future Data Centre technology
Azure Infrastructure Region
le 60 km
Data centres deployed in 50+ regions 10000s of kilometres of fibre
10-20 data centres region
~100000s servers data centre3
How do we scale cloud capacity
hellip
Faster servers
CPU DisksMem
GPUFPGA
1001101010
2013
1001101010
2020
44 ZB 44 ZB
0100010101
1101101010
1011000110
0100010101
1001101010
0010011011
1011101010
0001001110
0110001011
Source IDC 2014
Data doubling every two years
Need for low latency at scaleML and AI driving this further
Slide credit Mark Russinovich
How big is one Zettabytebull 1 ZB = 1021 bytes = 1 Billion TB
bull 10 TB HDD
bull Need 100 Million of those
bull Thatrsquos a 2500 km high tower
bull Data also needs to be transmitted across the network
bull 100 Gbps data centre network link
bull Thatrsquos 2500 years for transmission
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Hyperscale cloud hardware design in
collaboration with the Open
Compute Project (OCP)
Pushing compute
growth
Project Olympus Rack SystemProject Olympus
Custom storage racks optimised for
cold storage
Pushing storage growth
Disk dPower
for d
Cooling
for d
Project Pelican
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Azure Infrastructure Region
le 60 km
Data centres deployed in 50+ regions 10000s of kilometres of fibre
10-20 data centres region
~100000s servers data centre3
How do we scale cloud capacity
hellip
Faster servers
CPU DisksMem
GPUFPGA
1001101010
2013
1001101010
2020
44 ZB 44 ZB
0100010101
1101101010
1011000110
0100010101
1001101010
0010011011
1011101010
0001001110
0110001011
Source IDC 2014
Data doubling every two years
Need for low latency at scaleML and AI driving this further
Slide credit Mark Russinovich
How big is one Zettabytebull 1 ZB = 1021 bytes = 1 Billion TB
bull 10 TB HDD
bull Need 100 Million of those
bull Thatrsquos a 2500 km high tower
bull Data also needs to be transmitted across the network
bull 100 Gbps data centre network link
bull Thatrsquos 2500 years for transmission
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Hyperscale cloud hardware design in
collaboration with the Open
Compute Project (OCP)
Pushing compute
growth
Project Olympus Rack SystemProject Olympus
Custom storage racks optimised for
cold storage
Pushing storage growth
Disk dPower
for d
Cooling
for d
Project Pelican
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
How do we scale cloud capacity
hellip
Faster servers
CPU DisksMem
GPUFPGA
1001101010
2013
1001101010
2020
44 ZB 44 ZB
0100010101
1101101010
1011000110
0100010101
1001101010
0010011011
1011101010
0001001110
0110001011
Source IDC 2014
Data doubling every two years
Need for low latency at scaleML and AI driving this further
Slide credit Mark Russinovich
How big is one Zettabytebull 1 ZB = 1021 bytes = 1 Billion TB
bull 10 TB HDD
bull Need 100 Million of those
bull Thatrsquos a 2500 km high tower
bull Data also needs to be transmitted across the network
bull 100 Gbps data centre network link
bull Thatrsquos 2500 years for transmission
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Hyperscale cloud hardware design in
collaboration with the Open
Compute Project (OCP)
Pushing compute
growth
Project Olympus Rack SystemProject Olympus
Custom storage racks optimised for
cold storage
Pushing storage growth
Disk dPower
for d
Cooling
for d
Project Pelican
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
1001101010
2013
1001101010
2020
44 ZB 44 ZB
0100010101
1101101010
1011000110
0100010101
1001101010
0010011011
1011101010
0001001110
0110001011
Source IDC 2014
Data doubling every two years
Need for low latency at scaleML and AI driving this further
Slide credit Mark Russinovich
How big is one Zettabytebull 1 ZB = 1021 bytes = 1 Billion TB
bull 10 TB HDD
bull Need 100 Million of those
bull Thatrsquos a 2500 km high tower
bull Data also needs to be transmitted across the network
bull 100 Gbps data centre network link
bull Thatrsquos 2500 years for transmission
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Hyperscale cloud hardware design in
collaboration with the Open
Compute Project (OCP)
Pushing compute
growth
Project Olympus Rack SystemProject Olympus
Custom storage racks optimised for
cold storage
Pushing storage growth
Disk dPower
for d
Cooling
for d
Project Pelican
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
How big is one Zettabytebull 1 ZB = 1021 bytes = 1 Billion TB
bull 10 TB HDD
bull Need 100 Million of those
bull Thatrsquos a 2500 km high tower
bull Data also needs to be transmitted across the network
bull 100 Gbps data centre network link
bull Thatrsquos 2500 years for transmission
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Hyperscale cloud hardware design in
collaboration with the Open
Compute Project (OCP)
Pushing compute
growth
Project Olympus Rack SystemProject Olympus
Custom storage racks optimised for
cold storage
Pushing storage growth
Disk dPower
for d
Cooling
for d
Project Pelican
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Hyperscale cloud hardware design in
collaboration with the Open
Compute Project (OCP)
Pushing compute
growth
Project Olympus Rack SystemProject Olympus
Custom storage racks optimised for
cold storage
Pushing storage growth
Disk dPower
for d
Cooling
for d
Project Pelican
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Hyperscale cloud hardware design in
collaboration with the Open
Compute Project (OCP)
Pushing compute
growth
Project Olympus Rack SystemProject Olympus
Custom storage racks optimised for
cold storage
Pushing storage growth
Disk dPower
for d
Cooling
for d
Project Pelican
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Custom storage racks optimised for
cold storage
Pushing storage growth
Disk dPower
for d
Cooling
for d
Project Pelican
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Custom data centre interconnect
technology at multi-terabit speeds
Pushing networking
growth
Project Madison
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Pushing networking growth
11
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Driving exponential growth
Gb
ps
$
Network
FLO
PS$
Compute
GB$
Storage
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Disruption
NetworkCompute Storage
FLO
PS$
GB$
Gb
ps
$
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Disruption hellip and opportunitiesFL
OPS$
Compute
Gb
ps
$
Network
GB$
Storage
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Cutting-edge technologies in academic labs across the world
Southampton University University College London Cambridge University
Eindhoven University of Technology (Tue) and University of California Santa Barbara (UCSB)
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
A big opportunity hellip cloud-first approach
Optics
NetworkingStorage
Hardware
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Innovation across the cloud stackwwwopticsforthecloudcom
Custom switch
Storage on glassCustom optics chips
Physical
Network
Storage
New optics
New hardwaresoftware
New media
Application New scenarios
Optics for the Cloud
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Microsoft Future Decoded
Optical Network
for the Cloud
02112018
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Data Center Networks Today
19
Racks of servers interconnected by electrical packet switches
~100000 servers
1011hellip
3) Packets are queued
on output port
1011hellip
2) Address lookup
1) Optical-to-electrical
conversion
4) Electrical-to-optical
conversion
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Moorersquos Law for Networking
20
Gb
ps
$
Time
Electrical
switching
Double the speed every two years
at the same cost
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Free switch scaling enabled keeping
network cost low and constant
Clo
ud
netw
ork
co
st
Time
21
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Disruptionhellipand opportunities
22
Gb
ps
$
Time
Electrical
switching
Clo
ud
netw
ork
co
st
Time
Optical
switching
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Why Optics
23
1011hellip
1011hellip
Optical Fiber
1 Already used for transmission
2 Non-CMOS basedhellip
Optical Switch
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
How can we switch packets using light
24
1) Arrayed Waveguide Grating Router (AWGR)
passive component (no transistors or buffers)
3) Fast tunable laser
2) Wavelength-based forwarding
Port 1
Port 2
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
How fast can we tune the laser
25
Tunin
g t
ime
~1ms
10000x
~100ns
256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
Utilizationlt30
Need for a fundamentally
different approach
20 ns
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Microsoft Custom Tunable Laser
26
6mm
PIC Photonic Integrated Circuit
(Indium-phosphide)
Wavelength
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Fast Tunable Laser
27
Tunin
g t
ime
~1ms
lt1ns
10000x
~100ns
100x256-byte packet
100 GbpsOff-the-shelf laser
+ custom algorithms
20 ns 100 ns 20 ns
MS custom laser
(tuning time lt1ns)
20 ns
Utilization lt30
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
All-optical Data Center Network
28
1 2 3 N
1011hellip 1011hellip
Clo
ud
netw
ork
co
st
Time
Electrical
Optical
Speed-of-light networking
No packet processing or buffering within the network core
eg row of racks 250 nanoseconds
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Bufferless blessing or curse
29
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
30
The lack of buffer requires tight synchronization among nodes
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
31
Developed a fully decentralized solution with nanosecond accuracy
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
32
Synchronization accuracy = 100s picoseconds
(less than 1 billionth of a second)Clocks are not synchronized
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Project Timeline
33
2016 2018 2022+
Production
lt1ns laser tuning
Nanosecond time synchronization
Scaling
Deployment hurdles
Integration
Reliability
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
34
CPU Data
GPUFPGA
1 2 3 N
hellip
1011hellip 1011hellip
Single pool of resources
Data Center As a Computer
Which new application scenarios can be enabled
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Project
Silica
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Vo
lum
e o
f d
ata
Data generated per year
Data volume is growinghellip
Years
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Vo
lum
e o
f d
ata
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Data generated per year
Data we can afford to store
Data volume is growinghellip but storage is expensive
Years
Vo
lum
e o
f d
ata
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Data storage tiers
Hot
Warm
Cold
Hot Data
Cold DataCapacity (logscale)
Archival storage
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GBMagnetic media
Short
Long
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Todayrsquos archival storage technologies
Low High Total Cost
($ GB)
Med
ia
Life
tim
e [Ye
ars
]
Tape
~10 TB cartridge
HDD
~14 TB disk
Optical disk
300 GB
Silica
Up to 360 TB
Magnetic media
Short
Long
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
bull Raw material is SiO2
(Quartz)
bull 10000s year lifetime
bull Write Once Read Many
(WORM)
bull Persistent no bit rot or
media decay
bull Resilient EMF-proof
resistant to heat
Archival storage in Glass
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Writing data into the glasshellip
Pulsed lasers rarr
Time
~100 fs
Femtosecond (fs) pulsed laser
120783 119943119956 = 120783120782minus12078312078711995611994211994011995211995111994111995600000000000000001 seconds
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Picosecond VS femtosecond laser writing
Picosecond (10 x10-12 s) laser
induces voids with external stress
Femtosecond (10 x 10-15 s) laser
induced small gratings in quartz glass
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
bull Laser beam is focused inside
the glass
Femto-second laser writing
bull At the focus the glass is modified
creating a lsquovoxelrsquo
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Voxel structure nanograting
20 nm20 nm
bull Multi-level encoding using shape and orientation of structure
Self-Organized Nanogratings in Glass Irradiated by Ultrashort Light PulsesYasuhiko Shimotsuma Peter G Kazansky Jiarong Qiu and Kazuoki HiraoPhys Rev Lett 91 (2003)
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Disaggregated Write-Read technology
Write Head
Glass library
Read Head
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Reading system
bull Source is an LED
bull Optical imaging setup
bull Parallel readoutSilica slab
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Polarization readout
bull Several images per layer
are acquired
bull Voxel traces are decoded
bull Multiple bits of information
per voxel
Processed image
Raw images
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos ostionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos postionet
aute plaut aliquiaspit vellam arcient
quatum nimus esti unt harchil
invenducia dolut porem ne volo qui
bull Ut accatem rerum ea is cum ilicae
lab ipiendanis quiae eos
Read head demoA track is a stack of layers
Glass slide with
many tracks (lsquodotsrsquo
to the naked eye)
Each lsquodotrsquo is a 3D array of voxels
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Reading head demo
Stacks
of voxels
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Optics Research inspired and designed for the Cloud
Thank You
httpopticsforthecloudcom
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Conclusions
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You
Session Feedback
Please rate this session in the Future Decoded app
Microsoft UK AI Research Report
Download the AI Report at httpakamsUKAIreport
Visit our Hands-on Labs on Level 3
Try technology out with on-demand labs and expert help
Go deep with Documentation
httpdocsmicrosoftcom
Things to do next
akamsIgniteTourLND
Thank You