SC24 Nvidia’s accelerators are among the many most energy hungry machines of their class, but the chips proceed to dominate the Green500 rating of essentially the most sustainable supercomputers on the planet.

Eight of the ten most power-efficient methods on the bi-annual list employed Nvidia elements, and of these 5 had been powered by the GPU large’s 1,000-watt Grace Hopper Superchip (GH200).

The elements, which meld a 72-core Grace CPU primarily based on Arm’s Neoverse V2 design and 480 GB of LPDDR5x reminiscence with an H100 GPU with 96 to 144 GB of HBM3 or HBM3e reminiscence, have grow to be fairly common within the HPC neighborhood.

On the newest Inexperienced 500 listing, the chip powers each the primary and second best methods — EuroHPC’s JEDI and the Romeo HPC Middle’s Romeo-2025 machines, which achieved 72.7 and 70.9 gigaFLOPS per watt within the Excessive-Efficiency Linpack benchmark — that is FP64, after all.

The 2 methods are practically an identical, having been constructed utilizing Eviden’s BullSequana XH3000 platform and using the identical GH200 accelerators. Nvidia’s GH200 additionally claims place 4, six, and 7 on the listing with the Isambard-AI Part 1 (68.8 gigaFLOPS/watt), Jupiter Exascale Transition Instrument (67.9 gigaFLOPS/watt), and Helios GPU (66.9 FLOPS/watt).

JUPITER Exascale Development Instrument

The Jupiter Exascale Improvement Instrument … Source | Picture by Forschungszentrum Jülich / Ralf-Uwe Limbach

In the meantime, Nvidia’s venerable H100 powers the fifth, eighth, and ninth best machines, together with the Capella, Henri, and HoreKa-Teal methods.

It’s uncertain Nvidia will retain its excessive rating on the Inexperienced 500. Its Grace-Blackwell superchips are already on the way within the type of the two.7-kilowatt GB200 and the 5.4-kilowatt GB200 NVL4.

The brand new merchandise don’t all the time ship extra compute energy per watt.

From the A100 in 2020 to the H100 in 2022, FP64 efficiency jumped roughly 3.5x. Nonetheless, in comparison with the 1,200-watt Blackwell, the 700-watt H100 is definitely sooner in FP64 matrix math. In reality, for FP64, the one enchancment is for vector math, the place the incoming chip boasts 32 % larger perf.

So whereas Nvidia enjoys excessive positions on the Green500 at the moment, AMD is not out of the sport simply but. In reality, the Home of Zen’s MI300A accelerated processing unit claimed the quantity three spot on the newest listing with the Adastra 2 system.

In the event you’re not acquainted, AMD’s MI300A was announced a bit underneath a 12 months in the past and fuses 24-CPU cores and 6 CDNA-3 GPU dies right into a single APU with as much as 128 GB of HBM3 reminiscence on board, and a configurable TDP of 550-760 watts. And, a minimum of on paper, the half already boasts 1.8x the HPC efficiency of the H100.

Constructed by HPE Cray utilizing EX255a blades – as used on the planet’s most powerful publicly recognized supercomputer –Adastra 2 managed 69 gigaFLOPS/watt of efficiency. It is not alone both. The tenth best machine is one other MI300A-based machine at Lawrence Livermore Nationwide Laboratory referred to as RZAdams, which managed 62.8 gigaFLOPS/watt.

Scaling up

All of those methods within the Green500’s high 10 are actually properly above the 50 gigaFLOPS/watt goal essential to attain an exaFLOP of compute in a 20-megawatt envelope. However because it seems sustaining these ranges of effectivity at scale is moderately tough.

Trying on the three best machines on the Green500, they’re all on the small aspect. JEDI is rated for simply 67 kilowatts of energy. For comparability, the Swiss Nationwide Supercomputing Centre’s Alps machine – essentially the most highly effective GH200 system on the Top500 – achieves 434 petaFLOPS within the HPL benchmark whereas consuming 7.1 megawatts, making it the 14th best machine at 61 gigaFLOPS per watt.

It is a comparable story for Adastra 2, which is even smaller than JEDI at 37 kilowatts. In the event you might preserve 69 gigaFLOPS per watt at scale, you’d solely want about 25.2 megawatts to match El Capitan’s 1.742 exaFLOPS of real-world efficiency. In actuality, El Capitan wanted practically 29.6 megawatts of energy to attain its record-breaking run. ®


Source link