AMD, Intel, and Nvidia all love to talk about their supercomputer wins. Winning a supercomputer contract doesn’t just represent a few thousand CPU or GPU sales — it gives the company in question unique bragging rights depending on the target workload, system speed, energy efficiency, and other demonstrated technology milestones. Winning a world government as a supercomputing customer is typically read (or at least marketed) as a demonstration of trust in a company’s long-term product roadmap.
Today, AMD and HP are talking up their new partnership with Lawrence Livermore National Laboratories to build El Capitan — an exascale-class supercomputer scheduled to be delivered by early 2023. Intel is scheduled to deliver a one exaflop system by 2021, so we’re seeing a significant leapfrog in total compute power in a relatively short period of time. The system will be shared by three separate facilities: LLNL, Sandia National Laboratories, and the Los Alamos National Laboratory. The machine will be used to simulate nuclear reactions to ensure safe management of the US nuclear stockpile as it ages. El Capitan will also “provide opportunities for researchers to explore new applications using emerging data-intensive workloads such as modeling, simulation, analytics, and AI to support future NNSA missions.”
The next-generation system will be based on AMD’s upcoming fourth-generation Zen architecture, codenamed Genoa, “next-generation” Radeon Instinct GPUs based on a new compute-optimized architecture, and a third-generation iteration of AMD’s Infinity Fabric. The Zen 3 architecture is expected in 2020, but AMD hasn’t given a date for Zen 4, beyond acknowledging that the architecture is in-design. If the company keeps to a yearly cadence we’d expect to see Zen 4 by 2021; if there’s a Zen 3+ step, we’d expect the chip to drop in 2022.
According to HP, the 2 exaflop target represents a 1.3x increase in achievable performance compared with what was forecast seven months ago. AMD isn’t talking about the exact number of CPUs and GPUs or any specs for the processors but did acknowledge that much of the horsepower would be provided by GPUs, not CPUs. That’s not at all unexpected; the adoption of GPU accelerators is largely behind the sudden surge in supercomputing horsepower over the last decade.
HP Cray, as one might expect, will be closely involved in this endeavor. HP will use a future version of its Shasta Compute Blade technology for El Capitan, along with Cray’s new Slingshot interconnect. Two additional specific projects the supercomputer will be used for, in addition to nuclear stockpile modeling, is a partnership with GlaxoSmithKline to try and accelerate cancer drug development timelines from six years to one, and research conducted in partnership with the National Cancer Institute to understand the mutation of RAS proteins linked to 30 percent of human cancers.
Overall, it’s a strong win for AMD and a sign that Epyc is being evaluated more broadly across the industry. Whether this will lead to more wins for AMD in AI and perhaps a new focus on that market segment is less clear. This announcement is a win for Radeon Instinct as well, but it would be nice to see AMD talking more about that side of the business.
AMD, HP Unveil 2-Exaflop Supercomputer With Epyc, Radeon Instinct - ExtremeTech