US Nabs Supercomputer Crown, But China Dominates the Top500 List | Tech News
This week’s announcement of a new list of the world’s 500 fastest supercomputers, unveiled at the International Supercomputing Conference (ISC 2018) in Frankfurt, Germany, has good news and bad news for the US.
The good news is that for the first time since November 2012, the fastest computer in the world is a US machine. The Summit computer at the Department of Energy’s Oak Ridge National Laboratory (ORNL) is now on top of the list, with a sustained theoretical performance of 122.3 petaflops on the High Performance Linpack test used to rank the Top500 list.
This machine, built by IBM, has 4,356 nodes, each one equipped with two, 22-core IBM Power 9 CPUs and six Nvidia Tesla V100 GPUs, all connected by a Mellanox EDR InfiniBand network.
The No. 3 system is also new, the Sierra Machine at DOE’s Lawrence Livermore National Laboratory (LLNR), which delivered 71.6 petaflops on HPL. Built by IBM, Sierra’s architecture is quite similar to that of Summit, with each of its 4,320 nodes powered by two Power9 CPUs plus four Nvidia Tesla V100 GPUs and using the same Mellanox EDR InfiniBand as the system interconnect.
Summit and Sierra are both part of the DOE program known as Collaboration of Oak Ridge, Argonne, and Lawrence Livermore national labs (CORAL), which is aimed at accelerating the development of high-performance computing and speeding the way toward exascale computing.
The No. 4 system is the Tianhe-2 (or Milky Way-2) machine at the National Supercomputer Center in Guangzhou, China. This machine replaced its 5-year-old Xeon Phi accelerators with custom-built Matrix-2000 coprocessors, which increased its performance from 33.9 petaflops to 61.4 petaflops. The earlier score made it the top machine on the Top500 list from June 2013 until November 2015, and it had been in second place prior to the release of the latest list.
The fifth-place machine is Japan’s AI Bridging Cloud Infrastructure (ABCI) at the National Institute of Advanced Industrial Science and Technology (AIST), with an HPL score of 19.9 petaflops. This machine is powered by 20-core Intel Xeon Gold processors along with Tesla V100 GPUs.
It’s interesting to note that in what is arguably a better test – the High Performance Conjugate Gradients (HPCG) benchmark, which in turn is made up of several other tests – Summit and Sierra take the top two spots, followed by the K computer at Japan’s RIKEN Advanced Institute for Computational Science (AICS), an older system based on Fujitsu’s SPARC64VIIfx.
In any case, the good news for the US is that it now has bragging rights for the fastest machine in the world.
The bad news is that the rest of the Top500 list is increasingly dominated by China, which now has 206 of the 500 fastest machines in the world. The US, by contrast, is down to 124 of those machines, in what may be its lowest-ever share. In overall performance, though, the US is now on top, with 37.8 percent of total performance, compared with 29.3 percent for China.
The top four systems are much faster than anything else on the list, and that leads to these numbers. Note that China took the lead in the number of machines on the list six months ago, replacing the US, which has had the most number of systems on the list in the 25 years these systems have been tracked.
It’s likely that as a result, Chinese companies will now top the list of supercomputer vendors. Lenovo leads with a 23.4 percent market share, followed by HPE, Inspur, Sugon, and Cray. Lenovo is a Chinese-founded company with headquarters in Beijing and North Carolina; Inspur is a rapidly growing Chinese server maker, and Sugon is a Chinese firm that focuses on supercomputers.
On the Green 500 list, which focuses on power efficiency, the top three spots are held by Japanese systems, starting with the Shobusystem B; this is unchanged from the previous list. All three systems use Xeon main processors and PEZY-SC2 accelerators. Nvidia’s own DGX Saturn V Volta is in fourth place, followed by the Summit system.
The overall lists give something for many vendors to be happy about. IBM, for example, built the two new US systems.
Nvidia points out that its GPU accelerators – which the company now calls “Tensor Core GPUs” – are in five of the top seven systems (all except the Chinese systems); and 17 of the top 20 on the Green 500 list (all except those running Pezy accelerators). It said Summit has 27,648 GPUs, the most of any computer in the world, and that these provide 95 percent of its power. Nvidia also said Summit is capable of 3 exaops (3 billion billion lower-precision operations per second).
Meanwhile, Intel said its processors power 95 percent of all the systems on the Top500 list, and 97 percent of all the new systems, including 37 machines that use its new Xeon Scalable processors. Intel also said the next version of its Omni-Path Architecture, capable of connecting machines at up to 200 Gb/s, will be out next year.