Supercomputing 17 Features Some Surprises, Hints for the Future

In that location were a few surprises at this month'due south Supercomputing 17 conference, including a very good showing for the PEZY-SC2 accelerator, too as Intel'due south alter in management for its Xeon Phi line. In addition, we also saw the first systems built on Intel'southward Xeon Scalable Processor (Skylake) and the first to characteristic Nvidia's Volta accelerator.

But perhaps most notable is the news that Chinese systems at present account for 202 of the 500 supercomputers on the newest Meridian 500 listing compared with only 143 from the US American systems have dominated the list since its inception 25 years ago, and merely a few months agone, the US had 169 systems to China's 160. The same is true when it comes to amass performance, with the Chinese systems combined bookkeeping for 35.four percent of the total performance of the Top 500 systems.

The fastest computers in the world go along to be the two massive Chinese machines that take topped the list for several years now: Sunway TaihuLight, from Mainland china'south National Supercomputing Center in Wuxi, with sustained Linpack operation of more than 93 petaflops (93 thousand trillion floating point operations per second), and the Tianhe-two, from China's National Super Reckoner Center in Guangzhou, with sustained performance of more than than 33.viii petaflops. These remain the fastest machines, and by a huge margin. The Piz Daint arrangement from the Swiss National Supercomputing Middle, a Cray system that uses Intel Xeons and Nvidia Tesla P100s, held third identify with Linpack sustained performance of 19.6 petaflops.

The biggest change at the top is a new organisation in quaternary place: an upgraded version of the Gyoukou supercomputer, a ZettaScaler-2.2 system deployed at Japan's Agency for Marine-Earth Science and Applied science. This machine uses PEZY-SC2 accelerators, a second-generation 2048-core chip that provides a peak performance of 4.096 teraflops in double-precision, as well as conventional Intel Xeon processors, for a full of 19,860,000 cores. (An earlier ZettaScaler machine with the PEZY-SC2 made the list at a lower level on the June version). That gives it the highest number of cores used together—likewise known every bit the highest level of concurrency—seen to date, surpassing the TaihuLight, which has 10.half dozen million cores. The Gyoukou machine achieved a Linpack sustained operation of xix.14 petaflops, but what's interesting is that it uses ane.35 megawatts of ability, compared to ii.27 megawatts for PizDaint, 17.eight megawatts for Tinahe-2, and 15.4 megawatts for Taihulight. That's still a lot of power, but information technology'southward a big reduction compared to the other systems and a potent indication that power concerns are important, even for the fastest machines in the world. Information technology'south likewise worth noting that this shows how new architectures can reduce power depict dramatically.

The peak The states system remains the Titan supercomputer at the Oak Ridge National Laboratory, a five-year-former system that uses Nvidia K20x GPU accelerators and delivers 17.59 petaflops, which is at present in 5th identify on the list.

In the latest Green 500 list of the nearly energy-efficient supercomputers, four of the top 5 slots—including the top iii—went to newly-installed Japanese systems, all based on the ZettaScaler-2.2 compages and the PEZY-SC2 accelerator. The most efficient of these is the Shoubu System B, which is installed at RIKEN's Advanced Heart for Computing and Communication. The Shoubu System B accomplished 17.0 gigaflops/watt; Shoubu System B and the adjacent two systems, which use sixteen.8 and 16.7 gigaflops/watt respectively, are all in the bottom half of the Meridian 500 listing. The fifth system on the Green ranking is the Gyoukou system mentioned above—the number four organisation on the Height 500 list, at 14.ii gigaflops/watt.

These are big breakthroughs for the PEZY-SC2 accelerator, and may point to time to come directions for supercomputing architectures.

The quaternary greenest supercomputer is Nvidia'due south internal DGX SaturnV Volta system, which accomplished 15.i gigaflops/watt, and comes in at 149 on the Top 500 list. This arrangement has 22,440 Volta cores (which seem to be counted in a somewhat unlike way on the list than some of the other cores.) Nvidia has been having a very good twelvemonth for its accelerator, and has high hopes for more machines using the Volta GPU architecture.

As usual, the major vendors were crowing about their successes on the list, with Intel noting that its CPUs were in six of the top x systems and a tape loftier of 471 out of 500 systems. Intel besides noted that its new Xeon Scalable Processors were in 18 supercomputers with over 25 petaflops of performance. Only what may be more notable is that Intel said it is cancelling Knights Hill, the planned 10nm successor to the 14nm Knights Landing Xeon Phi processor. The visitor said it is now planning a new platform for exascale systems (one,000 petaflops) by 2022, but didn't divulge any details.

Nvidia emphasized that it had 34 new systems with its accelerators on the list, bringing the company's total to 87. Nvidia and partner IBM were crowing about the possibility that, past the fourth dimension the side by side list is due in June, the Summit machine at Oak Ridge National Laboratory (ORNL) should exist among the machines at the height of the list. This machine features 4600 nodes, each with ii IBM Power 9 base CPUs and 6 Nvidia Volta accelerators, with a projected performance of about 200 petaflops. This differs from Nvidia's internal solution, in that the CPUs and GPUs all communicate over NVLink 2.0 in a cache-coherent manner using OpenCAPI, so the GPUs tin can directly access primary system RAM. Acme will be followed by the Sierra machine at Lawrence Livermore National Laboratory, and past the AI Bridging Deject Infrastructure (ABCI) car in Japan.

I was interested to hear Cray announce a "production-set up" supercomputer based on the Cavium ThunderX2 processor on the 64-bit Armv8-A architecture, now available as part of its XC50 supercomputer. ARM-based machines are being tested at the Barcelona Supercomputer Center (whose Mare Nostrum automobile now ranks at number sixteen on the Top 500 listing), as well equally the "Mail-K" supercomputer in Nihon and the Isambard supercomputer in the UK. Cavium had some early on benchmarks on the ThunderX2, which showed the 14nm chip performing better on multi-threaded or memory-bound applications than Intel'due south Skylake Xeons, though Intel remains the leader in single-threaded and in raw teraflops. Annotation that in addition to the Cavium design, Qualcomm has also announced an ARM-based server bit chosen Centriq.

In other processor news, AMD announced it had ramped production of its Epyc processors, though these aren't yet in any Pinnacle 500 systems, while NEC announced versions of its Vector Engine PCIe accelerator menu for its new SX-Aurora TSUBASA supercomputer series, which accept a particularly fast retention bandwidth.

On the interconnects side, Mellanox said that 77 percent of new systems on the Top 500 listing apply InfiniBand, while Intel touted recent successes with its Omni-Path Architecture, which is mostly used in its Xeon Scalable Processor (Skylake) systems. Meanwhile, a number of vendors are looking to Gen-Z, designed to be a lower-latency, memory-centric approach for very high-speed connections betwixt compute and memory/storage devices.

Also, it is interesting to note that for the first time, all of the machines on the Top 500 are at present running Linux.

Betwixt the strong first impression from PEZY-SC2, Intel'due south decision to dump Knights Mill for a new architecture, Nvidia's Volta, and new competition from AMD, ARM vendors, and NEC, now is an exciting fourth dimension in the earth of supercomputing. Adjacent year'southward lists should be quite interesting, equally we see which architectures actually perform, and which are well-nigh efficient, as many of the vendors and the supercomputer sites effort to position themselves in the race to produce an exascale (1000 petaflop) calculator with a sub-xx megawatt power draw.

HTML MODULE 4165

Source: https://sea.pcmag.com/feature/18447/supercomputing-17-features-some-surprises-hints-for-the-future

Posted by: martineztorandready.blogspot.com

Share this post