Nvidia H100-based Henri system checks AMD’s Green500 lead • The Register

SC22 There is a new energy-efficiency king atop this fall’s Green500 record of the world’s greenest supercomputers, and it is a tiny 31-kilowatt cluster powered by Nvidia’s H100 GPUs.

Developed by Lenovo for the Flatiron Institute in New York, the two-petaflop Henri system is the primary and solely system utilizing Nvidia’s Hopper GPU structure to make it into this fall’s Supercomputing charts.

The system itself is basically extra of an HPC cluster than a supercomputer, much like Frontier or LUMI. Primarily based on Lenovo’s ThinkSystem SR670 V2 server platform, every node pairs two 32-core Ice Lake Xeon Scalable processors with 4 of Nvidia’s 80GB H100 GPUs. With a complete of 5920 CPU and GPU cores, Henri is the second smallest system on the record.

However not like the Top500, which prioritizes pure efficiency, the Green500 weighs that efficiency in opposition to a system’s energy consumption, rating methods based mostly on what number of gigaflops they will squeeze out of every watt.

At simply over 65 gigaflops per watt, Henri managed to move Oak Ridge Nationwide Laboratory’s TDS check rig, the earlier champion of effectivity. And this even if Henri is barely hitting 37.6 % of his peak theoretical efficiency of 5.4 petaflops.

This means that there is nonetheless a major quantity of grunt on the desk, and if Flatiron or Lenovo can higher deal with the linear efficiency scaling, system effectivity ought to enhance over time as effectively.

Clues to Hopper’s effectivity

Evaluating Henri to the subsequent most effective system additionally utilizing Intel CPUs and Nvidia GPUs, the Atos THX.AB cluster, the H100-based system is about 59% extra environment friendly.

Whereas it is exhausting to say how a lot of this may be attributed to Nvidia’s new Hopper structure within the H100, it actually appears promising.

As we have seen with methods like Australia’s Setonix, GPU acceleration has a huge effect on effectivity versus CPU computation. Setonix’s CPU-only configuration ranked at 338 on this fall’s Green500, whereas its GPU-accelerated configuration landed at quantity 4.

That mentioned, a bigger system utilizing the identical {hardware} as Henri most likely will not scale linearly. As methods get bigger, it is fairly widespread for efficiency overhead to eat into general system effectivity. For instance, the full-size Frontier supercomputer at ORNL is 74 occasions the dimensions of the Frontier TDS system, however that is about 16% much less environment friendly.

One other unknown is how Nvidia’s H100s will fare when paired with quicker PCIe 5.0-equipped CPUs like Intel’s Sapphire Rapids.

AMD dominates Green500 as soon as once more

As typical, the Green500 has seen fairly a reshuffle this fall. However as we noticed with this spring’s rating, HPE’s AMD-based methods continued to energy the vast majority of the highest 10 most effective methods.

Frontier TDS, Adastra, Setonix – GPU, Dardel – GPU, Frontier and LUMI – now holding second to seventh place on the Green500 – are all powered by HPE’s Cray EX235a platform which pairs AMD’s 64-core Epyc 3 Milan CPUs with the MI250x GPU chipmaker intuition.

The quantity 10 Champollion system in France, based mostly on HPE’s Apollo server platform, additionally makes use of AMD Epyc processors however opts for Nvidia’s A100 GPUs over Intuition.

The one non-AMD methods to make the highest 10 had been Atos’ THX.AB in eighth place and MN-3 in ninth. Each methods are outfitted with Intel Xeon Scalable processors.

Yow will discover a full evaluation of the Green500 right here.

Change on the horizon

AMD’s dominance on the prime of the Green500 could not final lengthy. Subsequent-generation CPUs and GPUs from the likes of Intel, AMD, and Nvidia will make their means into methods over the subsequent few months.

We discovered this week that the Adastra system can be among the many first to implement AMD’s Epyc 4 Genoa CPUs. The chips promise a 14% IPC enhance with 50% extra cores, but it surely’s exhausting to inform if that will likely be sufficient to maintain AMD’s lead.

Intel HBM-stacked Xeon Scalable processors and Ponte Vecchio GPUs, now referred to as Xeon Max and Knowledge Middle GPU Max, are already arriving at Argonne Nationwide Labs for integration into the Aurora supercomputer. The Los Alamos Nationwide Lab’s (LANL) Crossroads machine will even use Intel’s Xeon Max processors.

Equally, the primary supercomputers powered by Nvidia’s Grace and Grace-Hopper superchips, together with LANL’s Venado system, are anticipated to launch subsequent yr.

Relying on how effectively these chips carry out in HPC purposes, the Green500 might look very completely different subsequent spring. ®