Wednesday, November 20, 2013
Super Computing 2013 (SC13) has started in Denver, Colorado! The Super Computing top 500 list has already been updated, and there are no big changes to the top 5. Tianhe-2, a supercomputer developed in China, retained its position as the world’s No. 1 system.
One important but often ignored statistic for Super Computing is power consumption. For example the Tianhe-2 reaches 33.86 petaflop/s on the Linpack benchmark, but it consumes 17808 kW of power to do so! This is almost the entire output of a small thermal power station. The Green 500 ranks supercomputers by energy efficiency instead of just raw processing power. In terms of MFLOPS/Watt, the TSUBAME 2.5 developed by the Tokyo Institute of Technology should get their No. 1 spot. The latest Green 500 list will be available soon.
There are several techniques we can use to manage power consumption at the system level, but within each computing node we only have a few options. Heterogeneous setups, such as those using an NVIDIA/AMD GPU or Intel’s Xeon Phi have become the norm – and we think FPGA chips are another promising option. FPGA has pipeline parallelism, where several different tasks can be spawned in a push-pull configuration and each task has its own data supplied by the previous task – with or without host interaction. Several tech papers report that FPGA outperforms other chips in terms of both latency and energy eﬃciency. The standard FPGA languages (VHDL and Verilog-HDL) can be a complicated struggle for any software developer (like me!), but the chip vendor ALTERA has recently announced OpenCL support on their FPGA devices. This will make development faster, and cause fewer headaches. Fixstars has been a part of their early access program for their OpenCL SDK since last summer, and we provide OpenCL software and services to our clients.
Unfortunately, we don’t have a booth at SC13 this year, but we can be found at our partner Nallatech’s booth (#3519). We’ll even have coupons for copies of our OpenCL programming book.
Stop by and check out OpenCL on FPGA!