in-memory computing

23
In-memory computing In-memory computing A solution to the Von A solution to the Von Neumann bottleneck Neumann bottleneck Sylvain EUDIER Union College, 2004 MSCS Candidate

Upload: cicada

Post on 28-Jan-2016

56 views

Category:

Documents


1 download

DESCRIPTION

Sylvain EUDIER Union College, 2004 MSCS Candidate. In-memory computing. A solution to the Von Neumann bottleneck. Plan. Introduction to a new architecture Different architectures The C-RAM Architecture: Implication / Application Performances Conclusion. Introduction. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: In-memory computing

In-memory computingIn-memory computing

A solution to the Von A solution to the Von Neumann bottleneckNeumann bottleneck

Sylvain EUDIERUnion College, 2004 MSCS Candidate

Page 2: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

PlanPlan

Introduction to a new architectureIntroduction to a new architecture Different architecturesDifferent architectures The C-RAM Architecture:The C-RAM Architecture:

Implication / ApplicationImplication / Application PerformancesPerformances

ConclusionConclusion

Page 3: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

IntroductionIntroduction

Von Neumann architectureVon Neumann architecture The situation (gap evolution)The situation (gap evolution) Some improvements were madeSome improvements were made

Can we avoid this bottleneck? Can we avoid this bottleneck? (graph)(graph)

Page 4: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Different ArchitecturesDifferent Architectures

Architectures and designsArchitectures and designs IRAM (IRAM (designdesign)) RAW (RAW (designdesign)) CRAM (CRAM (designdesign))

Page 5: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

The C-RAM (Computational The C-RAM (Computational RAM) architectureRAM) architecture

ApplicationsApplications

PerformancesPerformances

ImplicationsImplications New software designNew software design Energy consumptionEnergy consumption

Page 6: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

CRAM ApplicationsCRAM Applications

Image processing : Low-level Image processing : Low-level adjustments (brightness, average adjustments (brightness, average filter…)filter…)

Databases searches : Equivalence, Databases searches : Equivalence, Extremes, between limits…Extremes, between limits…

Multimedia Compression: MPEG Motion Multimedia Compression: MPEG Motion estimationestimation

Page 7: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Performances - ConfigsPerformances - Configs

CRAM 200Mhz; 32MB; 64K PE’s on a CRAM 200Mhz; 32MB; 64K PE’s on a Pentium 133Mhz (simulated)Pentium 133Mhz (simulated)

Pentium 133Mhz with 32 MB RamPentium 133Mhz with 32 MB Ram Sun SPARC Station 167 Mhz CPU with Sun SPARC Station 167 Mhz CPU with

64 MB64 MB

Page 8: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Performances – Basic opsPerformances – Basic ops

Ops complexity

Page 9: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Performances - ComparisonPerformances - Comparison

Page 10: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

New Software Design (Step New Software Design (Step 1)1)

Think Parallel (pseudo code)Think Parallel (pseudo code)

Page 11: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

New Software Design (Step New Software Design (Step 2)2)

Use a different language (modified C++)Use a different language (modified C++)

Page 12: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

New Software Design (Step New Software Design (Step 3)3)

Possibly coding in assembly to optimizePossibly coding in assembly to optimize

Page 13: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

CRAM Energy ConsumptionCRAM Energy Consumption

We avoid the use of a busWe avoid the use of a bus

We have a direct access to memoryWe have a direct access to memory No overhead in communicationNo overhead in communication

Finally the CRAM use 20 times less Finally the CRAM use 20 times less energy therefore less heatenergy therefore less heat

Page 14: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

The future…The future…

Which architecture will be chosen?Which architecture will be chosen?

End of today’s architecture?End of today’s architecture?

A PetaOps is feasible with CRAMA PetaOps is feasible with CRAM

Blue Gene/P aims at the petaFlops Blue Gene/P aims at the petaFlops (view)(view)

Page 15: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Questions ?Questions ?

Page 16: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Memory bandwidth in a Memory bandwidth in a computercomputer

Back

Page 17: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

IRAM DesignIRAM Design

Back

Page 18: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

RAW DesignRAW Design

Back

Page 19: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

CRAM DesignCRAM Design

Back

Page 20: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Operations Complexity for Operations Complexity for CRAMCRAM

Back

Page 21: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Blue Gene/PBlue Gene/P

Scale

Page 22: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Computing power ScaleComputing power Scale

Back

Page 23: In-memory computing

Seminar - Processing in MemorySeminar - Processing in Memory

Memory – processors gap Memory – processors gap evolutionevolution

Back