7
 Peaceful coexistence among architectures Andreas Olofsson, Adapteva [email protected] HPEC 2009

Peaceful coexistence among architectures

Embed Size (px)

Citation preview

Page 1: Peaceful coexistence among architectures

  Peaceful coexistence among

architectures

Andreas Olofsson, [email protected]

HPEC 2009

Page 2: Peaceful coexistence among architectures

HPEC 20092

Introduction to Adapteva

We are a fabless semiconductor company busy developing the world's most energy efficient C­programmable floating point processors targeted at embedded applications.

Prototype in hand, product slated for middle of next year

Page 3: Peaceful coexistence among architectures

HPEC 20093

Why ASICs?

Because flexibility wastes energy:− Given the same conditions ASICs have a 50x­1000x energy 

efficiency edge over GPUs, DSPs, FPGAs, and Microprocessors.

The problems:− Expensive to (re)design, long turnaround times

− Very difficult to get specification right the first time

− Limited number of design teams who can design deep sub­micron ASICs

Page 4: Peaceful coexistence among architectures

HPEC 20094

Breaking down ASIC costs

"The cost of ASICs are sky rocketing", Moshe Gavrielov, CEO at Xilinx

SOC ASIC FPGAHARDWARE $16,875,000 $1,200,000 $900,000SOFTWARE $6,750,000 $300,000 $300,000TOOLS $5,000,000 $500,000 $0FABRICATION $2,500,000 $200,000 $0IP $5,000,000 $500,000 $0TOTAL $36,125,000 $2,700,000 $1,200,000

➢ Conclusion: Architecture development cost is heavily dependent on architecture and end application

Page 5: Peaceful coexistence among architectures

HPEC 20095

The “Ideal” Embedded Solution

Peaceful coexistence between architectures...

FPGA

Great at interfacesand bit level processing

Great at math!Limited flexibilitySimple programming

Great at control, human interfaces, and knowledge extraction

“SOFT”ASIC

Page 6: Peaceful coexistence among architectures

HPEC 2009

PIPELINEREGISTER FILE

DCACHE

CONTROL

ICACHE

6

Energy Waste in Processors

➢ Conclusion: The hardware­software energy efficiency gap can reduced from 100x to less than 3x with an appropriate architecture.

TYPICAL, 1% EFFICIENCY FOCUSED, 37% EFFICIENCY

REGISTER FILE

ARITHMETIC

PIPELINE

DCACHE

CONTROL

ICACHE

ARITHMETIC

GENERIC PROCESSOR

ALU

MEMORY CONTROL

Page 7: Peaceful coexistence among architectures

HPEC 20097

Adapteva's Vision for Soft ASICs

ANSI C­programmable IEEE Floating Point Signal Processing

16 to 4096 independent cores

50 GFLOP/Watt @ 65nmReal Performance

uP MEMORY

DMAROUTER