Friday, November 14, 2014
Data centric Coral supercomputer to reach 150 PFlops
The DoE announced today it has chosen IBM's "data centric" architecture for Oak Ridge National Lab's new "Summit" and Lawrence Livermore National Lab's new "Sierra," both to be completed by 2017.
Summit and Sierra will operate at 150 petaflops and 100 petaflops, respectively, compared to the world's current top supercomputer, the Tianhe-2 in China, which performs at 55 petaflops, Nvidia said in a separate news release. Summit will feature more than 3,400 nodes. The combination of very large memory per node and the powerful IBM POWER and NVIDIA processors provides an ideal platform for data analysis as well as computation.
Argonne National Labs will choose a different architecture (yet to be announced) for its pre-exascale supercomputer, reaching some 150 PFlops.
Perhaps the most interesting bit about these two computers, though, is their incredible power efficiency. Summit and Sierra are about five times more efficient than Titan, Sequoia, or Tianhe-2. 150 petaflops at 13.1 MW is about 11,000 megaflops per watt — way, way beyond any supercomputer currently in operation. Bluegene Q have had 1700 megaflops per watt efficiency, some 6.5 times less 6 years earlier.
To be precise. Then in 2022, IBM should be able to obtain some 55 GFlops per watt. That will be quite enough for EXAFLOPS level. Provided it will be able to calculate something else than the number of failures.
Anyway, 150 PFlops x 6.5 = 1020 PetaFlops under the same power consumption. Or even earlier to Exaflops goal at higher 20 MW power envelope.
99% BAD HARDWARE WEEK:
Summit will have a hybrid architecture, and each node will contain multiple IBM POWER9 CPUs and NVIDIA Volta GPUs all connected together with NVIDIA’s high-speed NVLink. Each node will have over half a terabyte of coherent memory (high bandwidth memory + DDR4) addressable by all CPUs and GPUs plus 800GB of non-volatile RAM that can be used as a burst buffer or as extended memory. To provide a high rate of I/O throughput, the nodes will be connected in a non-blocking fat-tree using a dual-rail Mellanox EDR InfiniBand interconnect.