future directions in reconfigurable computing · 2 2 | raztech llc outline brief history of...
TRANSCRIPT
Raztech LLC
Future Directions in
Reconfigurable Computing
Dr. Rahul Razdan
Contact:
2
Raztech LLC |2
Outline
Brief History of Electronics
Forward Looking Strategic Drivers
Current Situation in Reconfigurable Computing
Opportunities to Innovate
Conclusions
3
Raztech LLC |3
Accelerating Advances in Technology
Source: Kurzweil 1999 – Moravec 1998
Mechanical
Electro-Mechanical
VacuumTube
DiscreteTransistor
IntegratedCircuit
1900 1920 1940 1960 1980 2000
IE-5
IE-3
IE+0
IE+3
IE+6
IE+9
IE+12
$1000 B
uy
s: C
om
puta
tions
per
sec
ond
Nanotechnology
2020
4
Raztech LLC |4
Trends in Increasing Transistor Density
Source: Kurzweil 1999 – Moravec 1998
MaterialsScience
Era
NanoEra
LithoEra
Tra
nsi
stors
per
Chip
1985 1990 1995 2000 2005 2010 2015 2020198019751970
1 Million
1 Billion
1 Trillion
5
Raztech LLC |5
Architectural Progression (Technology)
Key Abstraction Innovations:
ISA, VM, Shared Memory
HLL Languages, operating systems
Mainline Innovations:
Latency Techniques:
Caching (istream, dstream), Prediction (istream mostly), Speculation
Bandwidth Techniques:
Pipelining, Wide Busses, Multi-processor/Multi-core
Smarter Compilers
RISC vs CISC, Performance Feedback
6
Raztech LLC |6
Architectural Progression (Technology)
Beyond Conventional CPUs:
Graphics, Network, Search?
Range of communication for multi-core/multi-processor (Ambric to Tilera)
RISC vs CISC of today.
Other Models: VLIW, Multithreaded, SIMD
Alternative Proposed Models:
Reconfigurable Computing Platforms
Nibble-mode processors
7
Raztech LLC |7
Business Overview:
Electronics: $1T in the context of $12T U.S. Economy ($60T+ WW)
Semiconductors: $200B
FPGA: $3.5B
EDA: $4B, PLM: $16B
Three Large Vertical Mega-Markets:
Information Technology: Computers, Laptops
Internet Infrastructure: Routers
Portable Consumer Device: cell phone, ipod
All else in noise:
Industrial, Medical, Defense, etc..
8
Raztech LLC |8
Outline
Brief History of Computer Architectures
Technology and Business View
Forward Looking Strategic Drivers
Technology and Business View
Current Situation in Reconfigurable Computing:
Opportunities to innovate:
Business and Technology View
Conclusions
9
Raztech LLC |9
1970 1975 1980 1985 1990 1995 2005 20102000
Increasing Cost (Technology)
Sources: IC Insights; Intel - Historical Wafer Size Trend
$0
$0.5
$1.0
$1.5
$2.0
$2.5
$3.0
Cost
per
fab
($bil
lions)
$3.5
$4.0POWER6790 million transistors
Intel 808629,000 transistorsSolid Logic
Tech
2 transistors
POWER1800,000 transistors
POWER4174 million transistors
Cell234 million transistors
10
Raztech LLC |10
IBM Transistor Performance Improvement
0%
20%
40%
60%
80%
100%
180nm 130nm 90nm 65nm 45nm 32nm
Gain by Traditional Scaling Gain by Innovation
Rel
ativ
e %
Im
pro
vem
ent
11
Raztech LLC |11
Semiconductor R&D Expense/Revenue
1
10
100
1000
10000
100000
1000000
1960 1966 1972 1978 1984 1990 1996 2002 2008
2004-20 annual revenue growth @ 6.5%, while R&D expense growth @ 12.2%
Revenue
Expense
$M
illi
on
s
Source: VLSI Research Inc.
12
Raztech LLC |12
Foundry industry evolution
Late 80s. Pure play foundry established to meet the need of manufacturing outsourcing from IDM
Mid. 90s. Fabless boomed with COT model, alternative of ASIC
Late 90s and early 00s. Fast growth of fabless attracted new foundries,
New pure play, Tower, Silterra, 1st Silicon, GSMC, SMIC
IDM turned foundries, Jazz, ASMC, HHNEC, Dongbu, VIS, Polar fab, Zfoundry
IDM offered foundry service, IBM, MagnaChip, Austriamicrosystems, Cypress
JV Fab with Fabless and Fab-lite, Chartered SMP, UMCi, Wafertech
Mid 00’, reality called, few foundries makes money
Differentiation -> Collaboration -> Consolidation started
Mega fab trend started
Samsung, Fujitsu, Toshiba in foundry game
13
Raztech LLC |13
Strategic Development Alliances: CAPEX Landscape
Source: IBM MI, IC Insights, August
2007
$33.22
$19.08
$12.65
$8.06
$4.83 $4.70 $0
$5
$10
$15
$20
$25
$30
$35
IBM Technology Alliances
Intel ASPLA TSMC UMC TI
2003-2006 Semiconductor CAPEX (non-DRAM), $B
14
Raztech LLC |14
Semiconductor Industry Association. International Technology Roadmap
for Semiconductors: 1999 edition. Austin, TX: International Sematech, 1999
Design Productivity GapL
ogic
Tra
nsi
sto
rs p
er C
hip
(M
)
1987 1989 1991 1993 1995 1997 1999 2001198519831981 2003 2005 2007 2009
0.001
10,000 100,000
0.01
Pro
duct
ivit
y (
K)
Tra
ns/
Sta
ff –
Mo
.
1,000 10,000
100 1,000
10 100
1.0 10
0.1 1.0
0.01 0.1
Logic Tr./Chip
Tr./S.M.
58% / Year Compounded
Complexity Growth Rate
21% / Year Compounded
Productivity Growth Rate
15
Raztech LLC |15
Impact on Design (Investment View)
Cost of IC design rising rapidly and very risky
$50M for 90nm
2-3 years before first revenue
7 years before exit event
Applicability to volume driven markets critical
single large focused market or
Collection of markets served with single solution
Increased focus on platforms for all vertical markets
16
Raztech LLC |16
What does this mean for computer architects ?
A small number of very large chips in traditional markets
Architectural innovation providing diminishing returns
Growth driven by GDP
Smaller Markets maybe big enough more platforms:
Bioinformatics, Cryptography, Oil Exploration, Search
Bets being made today: sicortex, pa-semi, pico-chip
Environment favors small number of flexible platforms
17
Raztech LLC |17
Outline
Brief History of Computer Architectures
Technology and Business View
Forward Looking Strategic Drivers
Technology and Business View
Current Situation Reconfigurable Computing:
Opportunities to innovate:
Business and Technology View
Conclusions
18
Raztech LLC |18
Reconfigurable Computing: Core Play Today
Assume a flexible implementation fabric (FPGA)
Operating principles:
Leverage flexibility for in-context optimization (performance/power)
Match data type and parallelism between implementation and application
Leverage space vs time
Various Levels of granularity:
ILP
Co-processor
System
Two Basic Use-Models: Embedded or HPC
19
Raztech LLC |19
Reconfigurable Computing: High Level Thrusts
ILP:
Research: PRISC, Chimeara, Garp
Commercial: ST (2007/2008)
Co-Processor:
Research: Splash
Commercial: Nallatech, Cray, HP
System:
Research: PAM, BEE2, Splash
Commercial: Emulation
20
Raztech LLC |20
Reconfigurable Computing: Observations:
HPC:
Many studies on “unit” level performance/power advantage
Limited Success at the Application Level
Limited Commercial Success
Embedded: some success with real-time data-flow applications
Radar Signal Processing
Image Processing
System: some success with dedicated system solutions (logic emulation)
21
Raztech LLC |21
Reconfigurable Computing: Logic Emulation
Two Basic Use Models:
Simulation Acceleration
In-Circuit Test
Historical:
Early Systems Built with FPGA Technology ($200M/year)
Issues With FPGA solution:
Multi-FPGA Mapping (Rent’s Rule)
Compile-Time limiting factor for design productivity
Today: nibble-mode uprocessors own largest marketshare
22
Raztech LLC |22
Reconfigurable Computing: Current Status
Limited Success: reasons cited
Design Flow is limited
FPGA support for runtime reconfiguration poor
Application entry point difficult
Observations:
Many Issues similar to parallel uprocessor development.
Caught between processor and ASIC
23
Raztech LLC |23
Outline
Brief History of Computer Architectures
Technology and Business View
Forward Looking Strategic Drivers
Technology and Business View
Current Situation in Reconfigurable Computing:
Opportunities to innovate:
Business and Technology View
Conclusions
24
Raztech LLC |24
Reconfigurable Computing: Potential Direction
Shift the value statement:
Focus on flexibility as a basic value (vs power/perf)
Find markets where it is inherently valued and build solutions
Long Life Cycle Markets(LLC):
Cost of deployment and maintenance overwhelm initial costs
Which markets exhibit these characteristics ?
Energy, Health Care, Aerospace and Defense
25
Raztech LLC |25
Opportunities to Innovate: LLC Products Details
Old Mega-Markets: Product Life (< 3 years)
New Potential Mega-Markets: Product Life (>10 years)
New Issues:
Future Requirements Volatility
Enhanced Reliability
Supply Chain Integrity
26
Raztech LLC
Short vs. Long Lifecycle
0 5 10 15
Product
Deployed
Product is
Replaced
Product is
Serviced
Product is
upgraded
Product
Deployed
Years
Short
Lifecycle
Long
Lifecycle
Product is
Serviced
Examples: DVD player, Cellphone, laptop
Examples: Pacemaker, Traffic control, power meter
27
Raztech LLC
Short vs. Long Lifecycle
0 5 10 15
Product
Deployed
Product is
Replaced
Product is
Serviced
Product is
upgraded
Product
Deployed
Years
Product is
Serviced
Time to market
Time in service
Short
Lifecycle
Long
Lifecycle
28
Raztech LLC
Short vs. Long Lifecycle
0 5 10 15
Product
Deployed
Product is
Replaced
Product is
Serviced
Product is
upgraded
Product
Deployed
Years
Product is
Serviced
Cost ofproduct
Cost ofmaintenance
Short
Lifecycle
Long
Lifecycle
Cost ofmaintenance
29
Raztech LLC
Short vs. Long Lifecycle
0 5 10 15
Product
Deployed
Product is
Replaced
Product is
Serviced
Product is
upgraded
Product
Deployed
Years
Product is
Serviced
Initialfunctionality
Evolvingfunctionality
Short
Lifecycle
Long
Lifecycle
30
Raztech LLC
Short vs. Long Lifecycle
0 5 10 15
Product
Deployed
Product is
Replaced
Product is
Serviced
Product is
upgraded
Product
Deployed
Years
Product is
Serviced
How soon will parts be available
How long will parts be available
Short
Lifecycle
Long
Lifecycle
31
Raztech LLC
Short vs. Long Lifecycle
0 5 10 15
Product
Deployed
Product is
Replaced
Product is
Serviced
Product is
upgraded
Product
Deployed
Years
Product is
Serviced
Missing features added to next version
Missing features added to upgrades
Short
Lifecycle
Long
Lifecycle
32
Raztech LLC
Long Lifecycle Solution: Programmable Fabric
Programmable Reconfigurable fabric can resolve the major LLC issues
Reliability
Evolvability
Supply Chain
All the years of research work on power/performance is “upside”
Potentially rich new area, but significant challenges exist.
33
Raztech LLC
Long Lifecycle Solution: Research Issues
LLC Design System:
What are the right metrics to capture the newer issues within the design environment ?
How are these metrics generated and maintained over time ?
Can we enhance optimization (synthesis) to account for the new metrics ?
How do we connect these metrics and view of lifetime cost to existing systems in the product lifecycle management space ?
LLC Programmable Platform:
How do we handle uprocessors, FPGAs, and analog ?
34
Raztech LLC
The Challenge : Traditional PLM Environment
ClearCase
SCM
ECAD
VPDM
AuthoringApplications
Tightly coupled
data management
with authoring
applicationsSystems
Model
RSA
SDM
Analysis &Simulation
SoftwareEngr. & Simulation
ECAD
PDM3
MCAD3CATIA
ENOVIA
OtherBusinessSystems
ERPParts
MgmtLogistics
Maint.
Systems
Legacy
Systems
PDM2
MCAD2
ePDM Enterprise PDM ECRECO
ActionActionAction
ECOECO
TightlyCoupled
Required
to maintain
integrity of
complex
product
models
Local DataManagement
35
Raztech LLC
Long Lifecycle Solution: Early Work
LLC Design System:
Early Research at University of Florida
Spun-out into startup:
Basic Idea:
Intercept the Design Process at the System Level (Systemc, Mathlab)
Introduce the notion of architecture and a mapping from functional definition.
Generate evolvability via process of sensitivity analysis of functional definition.
Generate reliability and supply chain from sensitivity analysis of architecture
Build a total cost function from the combination of above
4x-7x cost optimization shown in early prototypes
36
Raztech LLC
Long Lifecycle Solution: Early Work
Candidate
Algorithm
Model
Candidate
Architecture
Model
System
Model
Mapping
HW Analysis
PLM Code Generation
Algorithm
Model Library
Architecture
Model Library
37
Raztech LLC |
Conclusions:
Semi-Conductor Trends Point towards consolidation
Fabs and Designs
Reconfigurable Computing has had difficulty with performance/power
Design Eco-System
Caught between uprocessor and ASIC
Long Lifecycle Products offer an intriguing opportunity:
Flexibility itself is valued and can provide benefit
LLC Products may represent the next large markets (Energy/Medical)
Reconfigurable computing research maybe extendable in the LLC direction.