ISC ’22: The AMD-Intel-Nvidia HPC race heats up

The International Supercomputer Conference (ISC) kicked off in Hamburg, Germany this week with the release of the TOPP500 list of the fastest supercomputers, with a computer named Frontier taking first place.

Deployed at the US Department of Energy’s Oak Ridge National Labs, it is the first exascale machine (10¹⁸ floating point operations per second)—an HPE-Cray EX system powered by AMD Epyc CPUs and Instinct MI250 GPUs.

Intel had hoped to win the exascale battle with another DoE computer called Aurora, but AMD beat it to the punch. Frontier also beat out competitors from China and Japan that had hoped to win the exascale race.

The success of Frontier isn’t a surprise. A smaller version named Crusher was deployed at the Oak Ridge labs where it has been used to validate the applications to run on Frontier.

It’s yet another milestone in the comeback of AMD. Its server processors are now more than competitive with Intel’s best, and Instinct shows it can compete with coprocessors made by Nvidia, which has been its real nemesis. AMD has steadily gained ground on Intel and the server space, but Nvidia pretty much owns the show when it comes to AI and HPC.

AMD shouldn’t be popping the champagne just yet. Nvidia coprocessors can be found in 154 of the TOPP500 supercomputers; only seven of the supercomputers are using AMD Instinct cards.

Cray and AMD have one more thing to boast about: Frontier is also at the top of the Green500 list as the most energy-efficient supercomputer.

Intel updates its HPC roadmap

For some time, Intel has talked about its top HPC processor, Ponte Vecchio, the data center version of Intel’s Xe GPU architecture. Ponte Vecchio competes against Nvidia’s Ampere A100 and AMD’s Instinct MI2x0 chips, both of which are commercial GPUs repurposed and slightly altered for HPC and AI workloads.

Now Intel is moving beyond that processor and giving a glimpse into what’s coming after it: Rialto Bridge. Due next year, the new processor will provide around 30% better performance than Ponte Vecchio for applications, according to Jeff McVeigh, vice president and general manager of Intel’s Super Compute Group. (Can you say Osborne Effect)?

The successor to Rialto Bridge, Falcon Shores, is much more ambitious. It is a merging of both the Ponte Vecchio GPU and Sapphire Rapid-era of Xeon processors in a single package. That’s roughly what Nvidia is doing with its Grace Hopper Superchip, which consists of a 72-core CPU and a Hopper H100 GPU tightly connected by Nvidia’s high-speed NVLink-C2C chip-to-chip interconnect. AMD, your move.

McVeigh said Falcon Shores will provide five times better performance-per-watt, memory capacity, and memory bandwidth than “current platforms,” which likely means Nvidia’s A100 and the AMD Epyc server processor.

“Falcon Shores will bring together x86 and Xe architectures, taking advantage of next-generation packaging, memory, and I/O technologies giving huge performance and efficiency improvements for systems computing large data sets and training gigantic AI models,” he said in a video announcing the new chips.  

Nvidia Lines Up OEMs

Coming on the heels of the much-hyped introductions of its Grace CPU and Hopper GPU, Nvidia announced that Atos, Dell Technologies, GIGABYTE, HPE, Inspur, Lenovo, and Supermicro plan to offer servers built with both the Grace CPU and Grace Hopper Superchip. All these new systems will use the recently announced Grace and Grace Hopper reference designs which provide manufacturers the blueprints to rapidly prototype and build Nvidia-endorsed systems.

It also announced some future systems that could be candidates for the TOPP500 list. Los Alamos National Laboratory announced that its next-generation supercomputer, Venado, will be the first system in the US to be powered by Grace CPUs. It will be billed by Cray and use a mix of Grace CPU Superchip nodes and Grace Hopper Superchip nodes. When completed, its performance is expected to exceed 10 exaFLOPS.

Nvidia announced another HPE/Cray supercomputer, this one called Alps, located at the Swiss National Computing Center. Alps, built with the Grace CPU Superchips and intended for research in a range of fields, is expected to come online in 2023.