Nvidia’s H100 GPU, the flagship of its Hopper architecture, has debuted on the Top500 and Green500 lists at SC22. The new GPU appears in the relatively small Lenovo-built Henri system, which also features Intel’s Xeon “Ice Lake” CPUs. Henri is operated by the Flatiron Institute in New York City, a computational research institute focused on astrophysics, biology and quantum physics. Perhaps most notably, though, Henri is now the most energy-efficient publicly ranked supercomputer in the world, besting a bevy of Frontier-style systems that otherwise dominate the top ten of the Green500.
Henri is housed in a Secaucus, New Jersey, datacenter. For benchmarking, the Lenovo-built system leveraged ten of Lenovo’s SR670 V2 servers, each equipped with dual Intel Xeon “Ice Lake” CPUs, 1TB of memory and eight of Nvidia’s 80GB PCIe H100 GPUs (for a total of 80 H100s), all networked with Nvidia’s InfiniBand HDR. (After the benchmarking, the system swapped to a 20-node configuration with four H100s per node.)
Henri, which ranks 405th on the latest Top500 list, delivers 2.04 Linpack petaflops of computing power, relative to a theoretical peak of 5.42 petaflops. That underwhelming Linpack efficiency (37.6%), though, is easily counterbalanced by the machine’s stunning energy efficiency: Henri is capable of 65.09 gigaflops per watt, earning it the top spot on the Green500 list of the world’s most energy-efficient supercomputers.
This places the Intel- and Nvidia-based system comfortably ahead of the previous chart-topper, Frontier TDS (the one-rack test and development system for Frontier), which delivers 62.68 gigaflops per watt. HPE-built systems with that Frontier configuration – AMD Epyc “Milan” CPUs, AMD Instinct MI250X GPUs and Slingshot 11 networking – dominated the Green500 list in May, taking the top four spots: in order, Frontier TDS, Frontier itself, EuroHPC’s LUMI system and GENCI’s Adastra system. This year, the Frontier-style systems grew their share in the top ten of the Green500, adding the similarly-architectured GPU partitions of Pawsey’s Setonix system and KTH’s Dardel system.
Those systems – in order, Frontier TDS, Adastra, Setonix GPU, Dardel GPU, Frontier and LUMI – now occupy the 2nd through 7th places on the Green500 list. But in a coup, first place has been snatched by the H100-based system – a promising and surprising debut for Nvidia’s new chip after several of its top Green500 placements were snatched by those Frontier-style systems in June.
And that may not be the end of the story. With that 37.6% Linpack efficiency (relative to a 64.5% average Linpack efficiency on the November Top500 list), it seems likely that the benchmarking of Henri was a down-to-the-wire endeavor meant to debut the accelerators in time for SC22. “It looks like this was rushed so they could pass the test,” said Jack Dongarra in an interview with HPCwire. “Machines usually get between 70 and 85% of the peak, that’s respectable for HPL. So when you score so low, something’s not right. But it should leave room for further optimizations.” At the SC22 press briefing, Green500 list steward Wu Feng commented that Henri didn’t quite showcase the threefold energy efficiency improvement that was reported for the H100 vs. the previous-generation A100.
It’s easy to speculate, then, on whether Henri’s efficiency is something of a floor for an H100-based system – and, accordingly, whether we may see considerably more impressive efficiency benchmarking at ISC 2023. And, as HPCwire’s Tiffany Trader pointed out in her Top500 coverage, remember that this is the H100 when paired with Intel’s Ice Lake chips; even better numbers should follow pairing with (Intel’s) Sapphire Rapids or (AMD’s) Genoa chips. In a blog post, Nvidia promised further leaps in energy efficiency from technology like its Grace CPU Superchip, which the company says delivers “up to twice the performance per watt of a traditional CPU.”
Nvidia and Lenovo confirmed the crunched schedule for Henri in an email to HPCwire, calling the timeline “exceptionally fast,” adding that it was less than 36 hours from when the system was received until the power-on test, and noting that it was Lenovo’s first volume installation of the H100. Nvidia and Lenovo also affirmed that H100 deployments “will continue to improve flops/watt” as the companies work together on the hardware, elaborating that Hopper’s new power management abilities are dependent on variable OEM optimizations.
Of course, Henri isn’t just a technology showcase – it’s also a research supercomputer. The Flatiron Institute says that it will use the system – which joins two others at Flatiron (RE1, ranked 442nd on the Top500, and RE1A, ranked 488th) – for research in computational astrophysics, biology, mathematics, neuroscience and quantum physics.
“This supercomputer opens up opportunities for doing new kinds of science,” says Ian Fisk, co-director of the Flatiron Institute’s Scientific Computing Core. “This is a workhorse machine, and we’re going to let our researchers try new things and drive discoveries … This computer will allow us to do more science with smarter technology that uses less electricity and contributes to a more sustainable future. That’s what’s important to us.” (Further, the system is at least partially powered by renewable energy.)
In a touching note, Nvidia and Lenovo also told HPCwire that the privilege of naming the system went to the person who came up with the idea to benchmark it on the Green500 list. That person chose to name the system after his newborn son, Henri.
The Henri supercomputer became available to Flatiron researchers at the beginning of this month.