Fujitsu kicks off field trials for post-K exascale computing processor
Prototyping wrapped up with delivery around 2021
Field trials of Fujitsu's prototype exascale post-K supercomputer CPU have begun.
The post-K exascale (1,000 petaFLOPS) system project was started by Fujitsu and Japanese research institute RIKEN in October 2014. It followed the successful development of the K supercomputer, a 10.5 petaFLOPS system with 705,000 Sparc64 V111fx cores. The K project began in 2006 and the system was delivered in 2012.
Fujitsu and RIKEN expect to deliver an operational system around 2021.
The Arm8A-SVE (Scalable Vector Extension) 512bit CPU architecture has been used as the base for the post-K CPU. Fujitsu has worked with Arm to add half-precision arithmetic instructions, enhanced stacked memory bandwidth, and enhanced double-precision arithmetic performance.
HPE pulls sheets off largest Arm-based supercomputer AstraREAD MORE
The initial operation of this CPU has been verified. Fujitsu has released some post-K system details, saying there will be 48 cores plus 2 assistants for the computational nodes. There will also be 48 IO plus 2 assistant cores for the IO and computational nodes.
The interconnect will be a Tofu 6D Mesh/Torus. The system structure will have 1 CPU per node and 384 nodes per rack. We're not told the total number of nodes but expect it to be greater than 10,000.
Storage will be accessed through a FEFS (Lustre-based) global file system.
The power budget is set to be between 30 and 40MW.
RIKEN, Fujitsu and Arm are working on an Arm8 ecosystem. Software running on the K system should be recompilable to execute on the post-K. Visitors to ISC 2018 in Frankfurt, June 24-28, will be able to see the post-K CPU.
So far there are five progressing exascale system projects: three in the USA; Tianhe-3 in China; and the post-K in Japan. The EU has said it wants to build an exascale system but hasn't contracted with any supplier, such as Atos, to build one.
The three US systems are likely to involve Cray and Intel with an x86 CPU-based system, IBM with a POWER-based one, and HPE with either an AMD or an Arm-based system, with deliveries in the 2021-2023 period. Tianhe-3 could arrive in 2020. ®