Fujitsu doesn’t have a popularity as a serious microprocessor developer the way in which AMD, Intel, or ARM do, however the firm’s supercomputer and HPC work has resulted in a number of generations of high-performance techniques, traditionally constructed on the SPARC structure. Now the corporate is engaged on the successor to the Ok supercomputer it in-built partnership with RIKEN, the Japanese analysis institute. And it’s leaving SPARC behind for good in favor of a brand new, ARM-based structure.
The place the previous Ok supercomputer was constructed across the SPARC64 VIIIfx (itself primarily based on SPARC64 VII with quite a lot of HPC modifications), the brand new machine is a customized implementation of the ARMv8 instruction set. ARM, it’s possible you’ll recall, provides the choice to license both a particular ARM design just like the Cortex-A76 or an architectural license that enables an organization to construct a CPU of their very own design that maintains ARM compatibility. Fujitsu, like AMD, Apple, Intel, and Qualcomm, has chosen to go this route. And the corporate’s targets for the undertaking are extraordinarily aggressive, with a focused 100x efficiency enchancment over the Ok machine by 2021.
If this aim is realized, the post-Ok supercomputer can be one of many first machines to ship exascale compute efficiency and would signify a roughly 10x achieve over the present top-ranked supercomputer on the earth, the Summit system from IBM and Nvidia.
The brand new supercomputer will hit these heights partly through the use of ARM’s new 512-bit extensions for SIMD compute. In accordance with IEEE Spectrum, the brand new CPUs themselves include 48 main CPUs for computation and two assistant cores to deal with computational node visitors. There’s additionally a variant with 48 cores and four assistant cores to deal with I/O in addition to computational nodes, with one bodily CPU per node and 384 nodes per rack of area. Reminiscence bandwidth has additionally reportedly elevated by as much as an order of magnitude (for aggressive causes, Fujitsu and RIKEN aren’t disclosing what number of nodes the brand new system will likely be in whole). The brand new laptop additionally provides help for variable precision floating level to permit it to make use of diminished precision when applicable, and enhanced double-precision floating level efficiency in comparison with the sooner Ok machine. Nextplatform thinks each the 50-and 52-core CPU variants are prototypes as a result of presently low yields and that the completed product will possible goal for 64 cores on a 7nm course of.
Fujitsu is claiming the machine will solely want 30-40MW in contrast with 12.7MW for the Ok machine — a startling enhance in effectivity primarily based on superior energy administration strategies in addition to course of expertise enhancements. The machine may even embrace “excessive efficiency stacked reminiscence” (possible HBM3), and the core as a complete is alleged to supply “close to Xeon-Class efficiency.” All of this works out to an enchancment of 32-42x larger efficiency per watt in contrast with the Ok supercomputer. A part of the advance will likely be in course of expertise, for the reason that Ok laptop was constructed on 45nm, however there’s clearly plenty of extra work beneath the hood to make that form of enchancment potential.
Pulling off an exascale launch by 2021 can be one thing of an achievement on condition that US officers have been publicly cautious of such a guess — although it could nonetheless hold Horst Simon from having to kick out $2,000 to anybody who took his guess that we wouldn’t have exascale computing in-house by 2020.