what is next for accelerators? turf war or collaboration?
DESCRIPTION
What is next for accelerators? Turf war or collaboration? . Stefan Möhl, Co-Founder, Chief Strategy Officer, Mitrionics. Modern Hardware Accelerators. Modern Accelerators are COTS The main problem for HW accelerators in the past was the follow-on version - PowerPoint PPT PresentationTRANSCRIPT
What is next for accelerators? Turf war or collaboration?
Stefan Möhl, Co-Founder, Chief Strategy Officer,Mitrionics
Modern Hardware Accelerators• Modern Accelerators are COTS
– The main problem for HW accelerators in the past was the follow-on version
– COTS guarantees next-generation support
• COTS Accelerators Today:– FPGAs (The part we do)– GPGPUs– Many-Core CPUs (Intel, AMD, STI Cell)
Which Accelerator to Choose?FPGAs, GPGPUs or Many-Core?
GPGPUs• Advantages
– Very high off-chip bandwidth (80+ GB/s)– High performance on DP-FP (250 peak GFLOPS)– Good price/performance
• Disadvantages– Errors (lack of ECC and full IEEE-fp support)– Heat– Not x86 compatible
FPGAs• Advantages
– High off-chip bandwidth (20+ GB/s) 0.5 TB/s to on-chip memories
– HW acceleration on any operation Very high performance on non FP add/mul
– Low power (20 Watts)– Good price/performance
• Disadvantages– High unit cost– Not x86 compatible
Many-Core CPUs• Advantages
– Runs legacy code (at least on one core)– Runs full applications– Needed by other accelerators for Amdahl’s Law
A very fast device is needed to churn through the millions of lines of code that are not in the hot-spot!
• Disadvantages– Baseline performance– Heat– Runs legacy code slowly
Heterogeneous Computing
GPGPUsFloating Point
FPGAsNon-Floating Point
Many-core CPUsCommand & Control
Will the world go Heterogeneous?• The Big Guys have made the choice
• AMD + ATI = AMD Fusion– Heterogeneous chip with GPGPU + Many-core CPU– Intel are already the largest manufacturer of Graphics HW
• Intel Quick-path/AMD Hypertransport– In-socket FPGAs directly on the system bus– Possible to add Cache-coherency and TLB
• Accelerators have gone COTS– Will stay around, if used in HPC or not
• Legacy code must be fine-grain parallelized– For Many-Core, GPGPU and FPGA alike
Where does Mitrionics fit in?
The HPC FPGA Eco-SystemBioinformatics, Text search, Financials, etc
HW Module suppliers
System vendor
CPU/FPGA suppliers
MVP & SDK
App’s & Algo’s
The Mitrion Platform1) The Mitrion Virtual Processor
– A configurable processor design for a fine-grain massively parallel, soft-core processor
– 10-30 times faster than traditional CPUs at 20-25 Watts– Executes a program in an FPGA
2) The Mitrion-C programming language– An intrinsically parallel C-family language
3) The Mitrion Software Development Kit– Compiler– Debugger/Simulator– Processor configurator