Fujitsu announced today it will build the long-planned AI Bridging Cloud Infrastructure (ABCI) which is set to become the fastest supercomputer system in Japan and will begin operation in fiscal 2018 (starts in April). ABCI will use Intel’s Xeon Gold processors and Nvidia V100 GPUs and deliver 550 petaflops theoretical peak performance in half-precision floating point and 37 petaflops of double-precision peak floating point performance. The award is from Japan’s National Institute of Advanced Industrial Science and Technology (AIST).
The latest contract win means Fujitsu is now riding two CPU horses in the high stakes supercomputer race towards exascale. It is also building Japan’s post K supercomputer that is based on ARM processors. The post K machine, part of Japan’s Flagship 2020 Project, has encountered delays reportedly related to ARM development issues.
The new ABCI datacenter will be located on the Kashiwa II campus of the University of Tokyo. If this system had competed in the latest Top500 ranking of supercomputers published in June 2017, it would have taken the top position in Japan and third place globally. First reports roughly a year ago indicated the ABCI system target spec would be a 33-petaflops double-precision or 130-petaflops half-precision (see HPCwire article, Japan Plans Super-Efficient AI Supercomputer). The V100 tensor cores, which had not been announced when the plans became public, account for the much higher FP16 capability.
“The most noteworthy detail in the ABCI announcement is that it is being hailed – and configured – as a general-purpose supercomputer, not restricted to AI applications. The announcement highlights its double-precision performance, which is generally associated with scientific applications, as opposed to the single- or half-precision benchmarks that have come to be associated with deep learning,” said Addison Snell, CEO, Intersect360 Research.
“This win is also an important stepping stone for Fujitsu toward its Post-K architecture for exascale computing. Rather than SPARC processors, the ABCI system will use Intel Xeon processors with Nvidia Tesla GPU accelerators. Fujitsu can leverage this experience toward its eventual deployments that are ARM-based, with acceleration.”
The next Top500 list is due out at SC17 next month (Denver) and expectations are for shuffling at the top. In September, China’s released details of the upgrade to Tianhe-2 (MilkyWay-2) – now Tianhe-2A. It will use a proprietary accelerator (Matrix-2000), a proprietary network, and provide support for OpenMP and OpenCL. The upgrade is about 25 percent complete and expected to be fully functional by November 2017 according to a report by Jack Dongarra.
“The most significant enhancement to the system is the upgrade to the TianHe-2 nodes; the old Intel Xeon Phi Knights Corner (KNC) accelerators will be replaced with a proprietary accelerator called the Matrix-2000. In addition, the network has been enhanced, the memory increased, and the number of cabinets expanded. The completed system, when fully integrated with 4,981,760 cores and 3.4 PB of primary memory, will have a theoretical peak performance of 94.97 petaflops, which is roughly double the performance of the existing Tianhe-2 system. NUDT also developed the heterogeneous programming environment for the Matrix-20002 with support for OpenMP and OpenCL,” wrote Dongarra (Report on The TianHe-2A System).
It will be interesting to see if ABCI is stood up in time for next June’s Top500 list and where it lands. 37 petaflops (peak) should secure a top 10 or even top 5 placement but its enormous AI capability and low power will be the bigger story for many. According to today’s announcement, AIST has been planning to deploy ABCI as a global open innovation platform that will enable high speed AI processing by combining algorithms, big data and computational power. (Slide below taken from an early AIST presentation)
“As a cloud platform for AI applications offering the world’s top class machine learning processing capability, high performance computational capability, and energy efficiency, ABCI is expected to create new applications in a variety of fields. Furthermore, the system is foreseen to promote the utilization of cutting-edge AI technology by industry, including transfer of the latest cloud platform technology to the public through an open design,” said Fujitsu.
ABCI will feature a “high-performance computational system, a high-capacity storage system, and a variety of networking technology,” according to Fujitsu:
“[The core of ABCI] will consist of 1,088 PRIMERGY CX2570 M4 servers, mounted in Fujitsu’s PRIMERGY CX400 M4 multi-node servers. Each server will feature the latest components, including two Intel Xeon Gold processor CPUs (a total of 2,176 CPUs) and four NVIDIA Tesla V100 GPU computing cards (a total of 4,352 GPUs), as well as Intel SSD DC P4600 series based on an NVMe standard, as local storage.
“Moreover, the 2U size chassis PRIMERGY CX400 M4 can each mount two PRIMERGY CX2570 M4 server nodes with GPU computing cards, offering high installation density. In addition, by utilizing “hot water cooling” for its servers, this system can also realize significant power savings.”
Fujitsu has been investing heavily in AI and deep learning in recent years; that includes developing of a custom AI processor, the Deep Learning Unit (See HPCwire article, Fujitsu Continues HPC, AI Push). Fujitsu’s roadmap for the DLU includes multiple generations over time: a first-gen coprocessor is set to debut in 2018, followed by a second-gen embedded host CPU. More forward-looking are potential specialized processors targeting neuromorphic or combinatorial optimization applications. There was no mention of the DLU in today’s announcement.
Fujitsu says it plans to apply its AI and HPC technology to “AIST’s high system requirement standards for both hardware and software.” The company also plans to leverage lessons learned from the ABCI project to its Human Centric AI Zinrai initiative.