OHIO: IBM and Nvidia have been awarded $425 million by the US Department of Energy to build two brand-new supercomputers that leverage IBM’s Power 8 CPUs and Nvidia’s upcoming Volta GPUs. The two computers Summit, which will be built at Oak Ridge National Laboratory, and Sierra, built at Lawrence Livermore will have peak performance of around 150 petaflops when they’re completed in 2017-2018. This should make them the most powerful supercomputers in the world, though I wouldn’t be surprised if China is working on something even faster to retain its supercomputing crown.
The USA’s fastest supercomputers are Titan at ORNL and Sequoia at LLNL, both capable of around 17 petaflops, and placed second and third in the world respectively. Both of these supercomputers are actually quite new so I’m not sure if Summit and Sierra will replace them, or if all four supercomputers will exist at the same time.
Primarily there will be thousands of IBM Power8 CPUs running the show. Attached to those Power8 chips, via the new CAPI interface, will be thousands of Nvidia Volta GPUs. We don’t know a whole lot about Volta, except that it will feature stacked DRAM stacked on top of the GPU using through-silicon vias with memory bandwidth around 1TB/sec and very low latency. Nvidia’s current GPU accelerators are based on the Kepler core; Volta is the core after Maxwell.
Depending on their final specifications, IBM tells us that Summit and Sierra will use between 11.4MW and 13.1MW which, for systems capable of 150 petaflops, is very, very efficient. The world’s fastest supercomputer, Tianhe-2, uses 17 megawatts and maxes out at around 34 petaflops; the USA’s fastest supercomputer, Titan, hits 17.5 petaflops while using 8.2MW.
According to IBM, the secret sauce behind the massively improved efficiency of Summit and Sierra is “data centric computing,” which “puts computing power everywhere data resides.” When I pushed to find out what this actually means in terms of how the supercomputer will be laid out, no further data was forthcoming. Cutting through the PR fluff, it think it simply means that the Power8 architecture, combined with an InfiniBand interconnect, are capable of moving data around really quickly. (The press release mentions 17 petabytes per second.