Next-Gen AMD EPYC, Radeon Instinct Tapped For Supercomputer

AMD's promise of a 'fully optimized CPU and GPU design' for the word's 'most powerful supercomputer' underlines the new competition the chipmaker will face when Intel releases its first discrete GPUs for client and server systems next year.

AMD is apparently getting a leg-up over Intel in the high-performance computing space with the announcement that it will provide next-generation server processors for what CEO Lisa Su said will be the world's “most powerful supercomputer."

The Santa Clara, Calif.-based chipmaker announced on Tuesday that it will supply custom EPYC CPUs based on a future iteration of its Zen architecture, along with new Radeon Instinct GPUs and a custom Infinity Fabric interconnect, for a new supercomputer called Frontier.

[Related: Intel, AMD Lock Horns Over High-Performance Computing Prowess]

The new supercomputer, which is being developed for the U.S. Department of Energy at its Oak Ridge National Laboratory, will be capable of 1.5 exaflops of performance, or one-and-a-half quintillion floating-point computations per second. That would put Frontier above Aurora, a supercomputer Intel is developing for the Department of Energy that has promised 1 exaflop of performance. Summit, the world's current fastest supercomputer that uses IBM Power9 CPUs with Nvidia CPUs, is capable of performing 200 petaflops, or 200 quadrillion calculations per second.

The Frontier and Aurora supercomputers are set for completion in 2021.

For the Frontier project, AMD is partnering with supercomputing vendor Cray, which is also working with Intel on the Aurora supercomputer project. The total contract value of Frontier is more than $600 million, which puts it above the cost of the $500 million-plus contract value of Aurora.

In a press conference, Su said Frontier is "the first fully optimized CPU and GPU design optimized for artificial intelligence and supercomputing." AMD's CEO said the company will also supply a custom version of its Infinity Fabric interconnect, which will provide a "first of its kind" high-bandwidth, low-latency connection between the central processor and the GPUs in a one-to-four ratio design per node.

While Su didn't provide many details on the custom EPYC processor AMD is developing for Aurora, she said it will be based on a future iteration of its Zen architecture that comes after Zen 2, which AMD is using for its 7-nanometer EPYC "Rome" server processors coming out in the third quarter. The company has previously teased its Zen 3 architecture for a 2020 release window. Beyond hardware, AMD will provide an enhanced version of ROCm, its open-source programming platform for CPUs and GPUs.

"When you look at the amount of innovation that is going into Frontier […] we believe that at the time of power on the Frontier system really will be the most powerful super computer system in the world," she said.

Su's promise of a "fully optimized CPU and GPU design" underlines the new competition AMD will face when Intel releases its first discrete GPUs for client and server systems next year. Intel has stated that the Aurora supercomputer will use Xeon CPUs and the company's new "Xe Compute Architecture," which will address new workloads, particularly around artificial intelligence and data analytics.

Researchers at Oak Ridge National Laboratory plan to use Frontier's HPC and AI capabilities to "simulate, model and advance understanding of the interactions underlying the science of weather, sub-atomic structures, genomics and physics, and more," according to the Department of Energy.

"Frontier’s record-breaking performance will ensure our country’s ability to lead the world in science that improves the lives and economic prosperity of all Americans and the entire world," U.S. Secretary of Energy Rick Perry said in a statement. "Frontier will accelerate innovation in AI by giving American researchers world-class data and computing resources to ensure the next great inventions are made in the United States."