Nvidia exhibits off at Supercomputing 20

Practically 70% of the 500 fastests supercomputers on this planet as introduced on the Supercomputing 20 convention this week are powered by Nvidia, together with eight of the highest 10.

Amongst them was one named Selene that Nvidia constructed itself and that debuted at Quantity 5 on the semi-annual TOP500 checklist of the quickest machines. With top-end techniques requiring 10,000 or extra CPUs and GPUs, they’re enormously costly, so authorities or analysis establishments personal nearly all of them.

That makes Selene all of the extra uncommon. It was constructed by and is predicated at Nvidia’s Santa Clara, California, headquarters. (It’s extensively believed there are various supercomputers in non-public business that aren’t reported for aggressive causes.)

Nvidia’s Massive Exhibiting

Additionally vital is that one other Nvidia supercomputer, the DGX SuperPOD, took the highest spot on the GREEN500 checklist, which measures the vitality effectivity of the TOP500 techniques. 4 of the highest 5 techniques had Nvidia’s A100 Ampere GPU. Fujitsu’s Fugaku prototype, with simply Arm processors and no DRAM, fell from first place to sixth.

That is massive as a result of GPUs have by no means been identified for vitality effectivity however now Nvidia has a brand new story to inform: efficiency and vitality effectivity in a single product.

Nvidia additionally launched its Mellanox NDR 400Gbps InfiniBand household of interconnect merchandise, which shall be out there in Q2 of 2021. The brand new lineup contains adapters, knowledge processing items (DPUs), what Nvidia calls good NICs, switches and cables.

This isn’t only a doubling the bandwidth per port. Mellanox is tripling the variety of ports in a single gadget, which in principle will permit one change platform to attach all the knowledge middle. Mellanox stated adopters of NDR 400 Gbps InfiniBand can see a community price financial savings of 1.4x and energy financial savings of as much as 1.6x for datacenters.

AMD Claws Again

Excellent news and dangerous information for AMD. Its share of the highest supercomputers that use its CPUs practically doubled from 11 on the June TOP500 checklist to 21 on the present checklist. The expansion got here from new techniques with second-generation EPYC processors, which include an insane 64-cores.

On the down facet, it may possibly’t get any traction towards Nvidia on the GPU facet. Simply one of many high 500 used AMD Radeon GPUs. Even Intel’s Xeon Phi, which is discontinued, had a greater exhibiting with three techniques on the checklist.

However AMD is just not giving up. On Monday it revealed its new Intuition MI100 server GPU, calling it the “world’s quickest HPC accelerator for scientific analysis,” with greater than 10TFLOP for double-precision floating-point efficiency. AMD says it improves half-precision floating-point efficiency for AI coaching workloads by practically seven instances over the corporate’s earlier era of accelerators.

MI100 comes with a know-how known as Matrix Core, part of AMD’s new CDNA structure that’s designed for HPC and machine studying workloads. Future iterations of the structure shall be used for its next-generation Intuition GPUs.

Intel’s Newest Attempt at GPUs

Intel is hoping the third time would be the attraction for GPUs. It employed Raja Koudri, the designer of AMD’s Radeon GPU, to be its chief architect this time round so it actually has no excuse for technical failure.

Its new GPU known as the Xe, proving as soon as once more Intel has the worst product branding division within the Silicon Valley. The most important information concerning Xe was introduction of oneAPI Gold, the primary productized model of Intel’s programming platform for the Xe GPU line.

OneAPI Gold performs into Intel’s XPU technique of heterogeneous processing. Servers are way more than x86 chips. They’ve GPUs, FPGAs, AI accelerators, and community processors, and Intel has merchandise in each class. OneAPI Gold can rule all of them, permitting builders to write down one set of extremely optimized code and have it run optimally on any processor.

Intel is selling oneAPI as an open normal but it surely’s made for Intel’s structure. So I gained’t maintain my breath for AMD or Nvidia to undertake it any time quickly. However for anybody all-in with Intel, it might do what CUDA did for Nvidia.

Xe processors are nonetheless within the works, with the high-end model, codenamed Ponte Vecchio, due subsequent yr. OneAPI Gold is alleged to ship subsequent month.