“design in the late-silicon...

DUSD(Labs)

““Design in the LateDesign in the Late--Silicon Age”Silicon Age”

Jan M. RabaeyJan M. RabaeyUniversity of California @ BerkeleyUniversity of California @ Berkeley

Director MARCO Gigascale System Research CenterDirector MARCO Gigascale System Research Center

TriasicTriasic JurassicJurassic CretaceousCretaceous TertiaryTertiary QuarternaryQuarternaryPermianPermian

280M280M 12M12M

History Proceeds along AgesHistory Proceeds along Ages

Bronze Bronze Stone Stone IronIron

3500BC3500BC 1500BC1500BC2.5M2.5M

Old Old (Paleolithic)(Paleolithic)

MiddleMiddle((MesoliticMesolitic))

NewNew((NeoliticNeolitic))

Transitions between era’s most often are marked by mass extinctiTransitions between era’s most often are marked by mass extinctionsons

So Does the Electronics Age…So Does the Electronics Age…

How to avoid the mass extinctions How to avoid the mass extinctions that typically go with transitions between ages?that typically go with transitions between ages?

PrePre--siliconsilicon SiliconSilicon PostPost--siliconsilicon

1950’s1950’s Late 2010’sLate 2010’s

EarlyEarly--SiliconSilicon LateLate--SiliconSiliconSiliconSilicon

70’s70’s Late 90’sLate 90’s

The ASIC AgeThe ASIC AgeThe Custom EraThe Custom Era ????

ASICsASICs on the Road to Extinction (?)on the Road to Extinction (?)

0

2000

4000

6000

8000

10000

12000

1996 1997 1998 1999 2000 2001 2002* 2003* 2004* 2005* 2006*

ASIC ASSP

Courtesy R. Courtesy R. CamposanoCamposano, , SynopsysSynopsys

ASIC/ASSP Design StartsASIC/ASSP Design Starts

105105

9090

7575

6060

4545

3030

1515

2.52.5 3.53.5 55 1010 1515 2020

0.13nm0.13nm

Gate Counts in Millions of GatesGate Counts in Millions of Gates

Eng

inee

ring

Eng

inee

ring

Man

-Yea

rs

ComplexityComplexity

NRE: The Triple WhammyNRE: The Triple Whammy

Technology Node [nm]

02468

1012141618

250 180 150 130 100

Rel

ativ

e P

rice

Mask CostMask Cost

90 nm90 nm

and DSMand DSM

Source: Source: XilinxXilinx and and SynopsysSynopsys, Inc, Inc

An Era of Fewer, Flexible and Reusable ComponentsAn Era of Fewer, Flexible and Reusable Components

Platforms have taken a firm ground in application areas ranging Platforms have taken a firm ground in application areas ranging fromfromwireless, automotive, consumer, media processing, graphics and gwireless, automotive, consumer, media processing, graphics and gamingaming

XilinxXilinx Vertex ProVertex Pro

JanusJanus Chip Chip -- ST Micro and ParadesST Micro and Parades

Berkeley PleiadesBerkeley Pleiades

Managing Flexibility and ConcurrencyManaging Flexibility and Concurrency

Dramatically more SOC “programmers” than SOC “designers”Dramatically more SOC “programmers” than SOC “designers”

ManyOthersManyMany

OthersOthersRTOSRTOSRTOS

VHDLVHDLVHDL

System CSystem CSystem CCCC

The language soupThe language soup

Easier to create Easier to create heterogeneous concurrency heterogeneous concurrency than to use it!than to use it!

Need tools and frameworks that Need tools and frameworks that

support multiple Modelssupport multiple Models--ofof--

Computation in a seamless and Computation in a seamless and

expandable way.expandable way.

The Metropolis Environment (The Metropolis Environment (SangiovanniSangiovanni, UCB), UCB)

Watch Out for ParadigmWatch Out for Paradigm--Shifting SurprisesShifting Surprises

Structured ASIC orStructured ASIC or

“the return of gate arrays”“the return of gate arrays”

•• minimized mask costminimized mask cost•• manufacturablemanufacturable•• 50% area overhead for similar performance50% area overhead for similar performance•• example: VPGA (CMU)example: VPGA (CMU)

MasklessMaskless LithographyLithography

•• Eliminates mask costEliminates mask cost•• Puts some constraints on designPuts some constraints on design•• No to be expected any day, thoughNo to be expected any day, though

The Grand Challenges of the Late Silicon EraThe Grand Challenges of the Late Silicon Era

(1) Power and Energy(1) Power and EnergyPuts Bounds on Integration and Puts Bounds on Integration and PerformancePerformance

Limits the deployment of “truly Limits the deployment of “truly ambient electronics”ambient electronics”

Courtesy IBMCourtesy IBM

BerkeleyBerkeleyPicoNodePicoNode


(2) Uncertainty(2) UncertaintyProcess Variability puts Process Variability puts Synchronous Design Synchronous Design Paradigm under Severe StressParadigm under Severe Stress

Courtesy IntelCourtesy Intel


(3) Reliability(3) ReliabilityErrors Can and Will HappenErrors Can and Will Happen

Soft errors already a factSoft errors already a fact

Process scaling reduces Process scaling reduces SNRSNR

Some Bold Solutions…Some Bold Solutions…

(1) Aggressive Voltage Scaling(1) Aggressive Voltage ScalingPerformance through concurrencyPerformance through concurrency

Dynamic adaptation of supply and Dynamic adaptation of supply and threshold voltagesthreshold voltages

Careful orchestration of system activityCareful orchestration of system activity

NEEDS SYSTEM-LEVEL APPROACH!

> 1X Scaling ?> 1X Scaling ?MAC

ChandrakasanChandrakasan (MIT): (MIT): 175 mV DSP (ISSCC 02)175 mV DSP (ISSCC 02)


(2) Abandon the Purely (2) Abandon the Purely Synchronous ParadigmSynchronous ParadigmLate Binding Late Binding –– Let the system make Let the system make the timing decisionsthe timing decisions

Allow occasional timing errors to Allow occasional timing errors to happen and deal with them!happen and deal with them!


D-Cache

IF ID EX ME

M

WB

Register FileI-Cache

3 mm

3 mm

Razor (U. Razor (U. MichMich): Pseudo): Pseudo--synchronoussynchronous

Pleiades (Berkeley): Pleiades (Berkeley): Globally Asynchronous Globally Asynchronous –– Locally SynchronousLocally Synchronous


(3) Self(3) Self--Correcting ArchitecturesCorrecting ArchitecturesDesigns that detect and correct errorsDesigns that detect and correct errors

Careful use of redundancy and error Careful use of redundancy and error correctioncorrection

Provide reliable computation layered on Provide reliable computation layered on unun--reliable fabrics (as in the reliable fabrics (as in the communications world)communications world)


205 mm2

Alpha 21264

REMORAChecker

12 mm2

Self-checking processor (U. Mich)

StanfordStanford

Transitioning to the PostTransitioning to the Post--Silicon AgeSilicon Age

Implementation platforms that work under very low SNR, Implementation platforms that work under very low SNR, are nonare non--deterministic, unpredictable and unreliable…deterministic, unpredictable and unreliable…

MolecularMolecular

OrganicOrganic(polymer)(polymer) NanotubeNanotube

Daunting Perspectives? You bet!Daunting Perspectives? You bet!

The search for solutions needs the attention of the brightest miThe search for solutions needs the attention of the brightest minds all over the nds all over the

country and the world. Hence the country and the world. Hence the MARCO Focus Research CenterMARCO Focus Research Center Program!Program!

Bringing together in a collaborative setting the leading minds iBringing together in a collaborative setting the leading minds in n design technology for electronics systemsdesign technology for electronics systems

Daunting Perspectives? You bet!Daunting Perspectives? You bet!The search for solutions needs the attention of the brightest mThe search for solutions needs the attention of the brightest minds all over the inds all over the

country and the world. Hence the country and the world. Hence the MARCO Focus Research CenterMARCO Focus Research Center Program!Program!

““The design, verification, and test of complex, heterogeneous embedded

systems-on-a-chip/package, covering the complete spectrum from

system specification to implementationon emerging circuit fabrics.”

Gigascale System Research Center (GSRC)Gigascale System Research Center (GSRC)

SummarySummary

The semiconductor industry is The semiconductor industry is

facing a challenging time. facing a challenging time.

Radically new disruptive Radically new disruptive

solutions are needed.solutions are needed.

The GSRC collaborative model The GSRC collaborative model

to paradigmto paradigm--shifting research has shifting research has

proven to be very successful. proven to be very successful.

Not Just Research as UsualNot Just Research as Usual

Supporting the transition from Supporting the transition from

the late to the post silicon agethe late to the post silicon age

Avoiding MassAvoiding Mass--Extinctions Through Preventive ActionExtinctions Through Preventive Action

Thank you!Thank you!

http://http://w

ww.gigascale.org

www.gigascale.org

“design in the late-silicon...

Documents