enabling the a.i. era - applied materials materials_ai... · performance improvements over time...
TRANSCRIPT
External Use
Enabling the A.I. Era: From Materials to Systems
Sundeep Bajikar
Head of Corporate Strategy & Market Intelligence, Applied Materials
A.I. Summit at Semicon Korea | January 23, 2019
| External Use
Forward-Looking Statements
This presentation contains forward-looking statements, including those regarding
anticipated growth and trends in Applied’s businesses and markets, industry outlooks,
technology transitions, and other statements that are not historical facts. These statements
are subject to risks and uncertainties that could cause actual results to differ materially
from those expressed or implied by such statements and are not guarantees of future
performance. Information concerning these risks and uncertainties is contained in Applied’s
most recent Form 10-Q and other filings with the SEC. All forward-looking statements are
based on management's current estimates, projections and assumptions, and Applied
assumes no obligation to update them.
2
| External Use 3
Key Messages
* A.I. refers to Machine Learning and Deep Learning approaches.
A.I. Computing
Requires a New
Semiconductor
Playbook
A.I.* Requires
New Computing
Architectures
PART 1 PART 2
| External Use 4
Key Messages
A.I. Requires
New Computing
Architectures
PART 1
What is different
about A.I. computing?
Empirical analysis –
A.I. / IoT unaffordable
with current
computing model
| External Use
2017 2018E 2022E
HOME and BUILDINGS
HUMANS
INFLECTION YEAR Data generated by machines > humans
SMART FACTORIES
SMART VEHICLES
Humans
Machines
1.5ZB 2ZB
>10ZB
53% 44% <10%
SAFETY and SECURITY
SOURCE: Applied Materials model based on forecasts published by Cisco, Intel, Western Digital
5
Explosion of Data Generation
| External Use 6
Source: Computer Architecture: A Quantitative Approach, Sixth Edition, John Hennessy and David Patterson, December 2017, Applied Materials internal analysis
Deep Learning Enabled by Processing Abundant Data
What’s different about
A.I. computing?
Memory Bound
Performance limited by memory latency
and memory bandwidth
High Data Parallelism
Identical operations can be done on
multiple chunks of data in parallel
Low Precision
Training and inference require much
lower floating point (16 bit) and integer
(8 bit) precision
Relative to traditional
CPU architectures
Traditional multi-level on-die cache
architecture is unnecessary
Traditional deep instruction pipeline
and out-of-order execution are
unnecessary
Traditional 64 bit or higher precision
data paths are unnecessary
| External Use 7
Data Algorithms
ML / DL
RELATIONSHIP BETWEEN
Complexity of algorithms
VERSUS
Abundance of (training) data
+ affordable high-
performance computing
AI is about
“relentless classification”
Data quality and size are
making AI a reality
Source: Applied Materials internal analysis.
| External Use 8
ML / DL
CASE STUDY: Google Translate Circa 2010
► Complex algorithms based on
grammar and semantic rules
► Developed by team of linguists
70% accuracy
Source: Applied Materials internal analysis.
| External Use 9
ML / DL
CASE STUDY: Google Translate Today
► Simple algorithm
► Trained using every single web page
available in English and Chinese
► No Chinese speakers on team
98% accuracy
Source: Applied Materials internal analysis.
| External Use
Homogeneous / Discrete Compute Era
10
From Homogeneous to Heterogeneous Computing
CPU
Normal Compiled
/ Managed Code (Office, OS, Enterprise)
CPU
Client CC, Search,
Audio, VOIP
GPU
CPU
Imaging,
Video Playback
GPU
ASIC / FPGA
CPU
Gaming, HPC, Highly
Parallel Workloads
GPU
CPU
AI Workloads: Training, Inference,
Analytics, etc.
GPU
ASIC / FPGA
Heterogeneous Compute Era
CPU 0
CPU 2
CPU 1
CPU 3
CPU 4
GPU
Media
ISP
Audio
ASIC
FP
GA
Exploits data level parallelism
Uses the most energy-efficient
compute engine for a given job
Source: Applied Materials internal analysis.
| External Use
y = 0.0081x1.3065
R² = 0.9563
-
5
10
15
20
25
- 100 200 300 400 500
DR
AM
Sh
ipm
en
ts (
EB
)
Incremental Data Generation (EB)
11
In Current Compute Model…
y = 0.0022x2.0399
R² = 0.9569
-
100
200
300
400
500
- 100 200 300 400 500
NA
ND
Sh
ipm
en
ts (
EB
)
Incremental Data Generation (EB)
DRAM NAND
Source: Cisco VNI, Cisco, Gartner, Factset, Applied Materials internal analysis
2006 to 2016 data from Cisco and Gartner. 2017 to 2020 projections from Cisco for data generation (VNI IP traffic), and industry average estimates for DRAM and NAND content shipments
| External Use
y = 0.0081x1.3065 R² = 0.9563
-
20
40
60
80
100
120
140
160
180
- 500 1,000 1,500 2,000 2,500
DR
AM
Sh
ipm
en
ts (
EB
)
Incremental Data Generation (EB)
1% adoption
of L4/L5
Smart Car
y = 0.0022x2.0399
R² = 0.9569
-
2,000
4,000
6,000
8,000
10,000
12,000
14,000
- 500 1,000 1,500 2,000 2,500
NA
ND
Sh
ipm
en
ts (
EB
)
Incremental Data Generation (EB)
1% adoption
of L4/L5
Smart Car
12
A.I. / Big Data Era Needs New Compute Models
Source: Cisco VNI, Cisco, Gartner, Factset, Applied Materials internal analysis
2006 to 2016 data from Cisco and Gartner. 2017 to 2020 projections from Cisco for data generation (VNI IP traffic), and industry average estimates for DRAM and NAND content shipments
1% SMART CAR
ADOPTION
5X MORE DATA
8X MORE DRAM
IN 2020
1% SMART CAR
ADOPTION
5X MORE DATA
25X MORE NAND
IN 2020
DRAM NAND
| External Use 13
A.I. Needs New Energy-Efficient Compute Models
0%
5%
10%
15%
20%
25%
0
50
100
150
200
250
2016 2017 2018E 2019E 2020F 2021F 2022F 2023F 2024F 2025F
A.I
. a
s a
% o
f G
lob
al E
lec
tric
ity C
on
su
mp
tio
n (
%)
A.I
. TA
M (
$B
)
TR
AIN
ING
IN
FE
RE
NC
E
Source: Applied Materials internal analysis.
Estimated current
computing models
for A.I. represent 25%
of global energy
consumption by 2025
| External Use 14
Key Messages
A.I. Computing
Requires a New
Semiconductor
Playbook
PART 2
From classic 2D scaling
to architecture
innovation
From traditional
materials to materials
breakthroughs
| External Use 15
“Classic 2D Scaling” to Architecture Innovation
Source: Applied Materials internal analysis . Performance broadly refers to speed or power efficiency. ~2x refers to doubling of performance every two years consistent with Moore’s Law.
ISS Blog post #1: Nodes, Inter-nodes, Leading-Nodes and Trailing-Nodes
ERA
ORIGINAL
PERFORMANCE
PERFORMANCE
GAIN HOW
1x ~2x PC, Mobile
Classic 2D scaling When Moore’s Law scaling was
at it’s peak. Assumes no major
architecture changes
1x ~2x to
2,500x A.I. / IoT
Architecture Innovation Small benefit from Moore’s Law
scaling
| External Use 16
Litho-enabled Moore’s Law is Slowing
PERFORMANCE IMPROVEMENTS OVER TIME (VS. VAX-11/780)
100,000
10,000
1,000
100
10
19
78
19
86
20
03
20
11
20
15
20
18
25% per year
52% per year
23% per year
12% per year
3.5% per year
End of Dennard Scaling
Limits of parallelism of Amdahl’s Law
End of Moore’s Law
VAX-11/780
TIME BETWEEN LOGIC NODES IN YEARS
5.3
2.5
2
1.8
14 to 10nm
22 to 14nm
32 to 22nm
45 to 32nm
SOURCE: Computer Architecture: A Quantitative Approach, Sixth Edition, John Hennessy and David Patterson, December 2017 SOURCE: Bernstein
| External Use 17
A.I. Enabled by Materials Breakthroughs
Source: Applied Materials internal analysis . Performance broadly refers to speed or power efficiency. ~2x refers to doubling of performance every two years consistent with Moore’s Law
ISS Blog post #2: Nodes, Inter-nodes, Leading-Nodes and Trailing-Nodes.
ERA
ORIGINAL
PERFORMANCE
PERFORMANCE
GAIN
MATERIALS
ENGINEERING
1x ~2x PC, Mobile Unit Materials
1x ~2x to
2,500x A.I. / IoT
Materials Breakthroughs New devices, 3D techniques,
Advanced packaging
| External Use 18
Architecture Innovation
High Memory Bandwidth
Performance limited by memory
latency and memory bandwidth
High Data Parallelism
Identical operations can be done on
multiple chunks of data in parallel
Low Precision
Training and inference require
much lower floating point (16 bit)
and integer (8 bit) precision
enabled by
Materials Engineering
New Devices
On-die memory – e.g. MRAM
New memory – e.g. Intel® 3D XPoint™ technology
3D Techniques
Self-aligned structures – e.g. vias,
multi-color patterning
“Side-to-side” to “top-to-bottom”
– e.g. COAG
Advanced Packaging
Stacked DRAM – e.g. HBM / HBM2
Logic-DRAM
Sensor-Logic
Intel and 3D XPoint are trademarks of Intel Corporation or its subsidiaries in the U.S. and/or other countries.
| External Use
NEW PLAYBOOK NEEDED
Design + Manufacturing 19
ENABLED BY
Advanced packaging
New structures / 3D
New ways to shrink
New architectures
New materials PPAC
POWER
PERFORMANCE
AREA-COST
| External Use
Example: STT-MRAM
SAF PINNED LAYER
CAPPING
MGO CAP
FREE LAYER
TUNNEL BARRIER REFERENCE LAYER
SEED LAYER
New Memory Device Enabled By Materials Breakthrough
Source: Applied Materials internal analysis .
20
>15 individual layers of film
in MTJ stack
Each layer 0.2-to-2nm thin
Requires highly complex
process equipment
Requires materials
(Element, Composition,
Stack) and Etch co-
optimization
| External Use 21
Investments to Enable New Semiconductor Playbook
Maydan
Technology Center Applied Ventures META Center
Applied Packaging
Development Center
AI DESIGN FORUM
A. I. Design Forum
World’s most advanced
300mm R&D lab with
$300M invested in
past 3 years
>$250M assets
under management and
>30 active companies
Expands R&D capability
to support new types of
collaboration from
lab to fab and accelerate
innovation from materials
to systems
Announced November 2018
Scheduled to open in 2019
Fully-integrated 300mm
advanced WLP process
flow and a full suite of
equipment
17K sq. ft. cleanroom,
fully staffed
Expanded ecosystem
engagements
| External Use
A.I. – Big Data Era = the biggest
opportunity of
our lifetimes
Hardware renaissance
Materials innovation to enable new architectures, structures, ways to shrink and packaging approaches
New ecosystem playbook for Design and Manufacturing
UNLOCKED BY
22