With China’s announcement of TaihuLight, the title holder for world’s fastest supercomputer, the US is looking for its answer. And now it’s here….well, almost.
The Department of Energy’s (DOE) Oak Ridge National Laboratory is expecting the delivery of a new IBM system, named Summit, in early 2018 that will be capable of 200 peak petaflops.
Sunway’s TaihuLight can reach a theoretical peak speed of 124.5 petaflops, and has achieved 93 petaflops on the Linpack benchmark. Notably, for the first time ever, it uses no US-sourced parts at all, as it’s powered by Sunway 260-core SW26010 processors. That’s because of a ban of Intel’s high-end microprocessors to China’s four main supercomputing centers.
In contrast, Summit will employ IBM Power9 and Nvidia Volta GPUs. Summit, which has 3,400 nodes, will have each node delivering “over half a terabyte” of so-called coherent memory (HBM + DDR4), plus 800GB of non-volatile RAM that serves as a burst buffer or extended memory.
The DOE, which holds most U.S.government supercomputers used for scientific research, has two other major supercomputers planned for 2018.
One system, named Sierra, is a planned 150-petaflop IBM system that will be located at the Lawrence Livermore National Lab, and is scheduled to be available for use by mid-2018. A third supercomputer, a Cray and Intel system called Aurora, is due by late 2018, at the Argonne National Laboratory.
In a statement via Computerworld, the DOE pointed out that, since 1993, US supercomputing capabilities “have grown exponentially by a factor of 300,000,” and that “high-performance computing remains an integral priority for the DOE.”
While China may be record holder, that may not really mean much. Hardware speed and capability are important metrics, but equally important is the software package that will be run. “Strength of the U.S. program lies not just in hardware capability, but also in the ability to develop software that harnesses high-performance computing for real-world scientific and industrial applications.”