With its 200-petaflop peak performance, the Department of Energy’s Summit system sits comfortably in first place on the TOP500 ranking of the world’s most powerful supercomputers. Today, the agency revealed plans to build a new system that will pack seven times more computing power than its predecessor.

The DOE has awarded the $600 million contract for the project to supercomputer builder Cray Inc. and Advanced Micro Devices Inc., which will supply chips for the system. They expect to complete construction in 2021.

Dubbed Frontier, the supercomputer’s top speed will exceed 1.5 exaflops, which amounts to 1.5 million trillion computing operations per second. It will comprise of more than 100 server cabinets that are expected to take up an area the size of two basketball courts. These enclosures will be based on Cray’s newest Shasta architecture and will be fitted with servers likewise designed by the firm.

The compute nodes inside Frontier’s cabinets are set to be powered entirely by AMD silicon. Each server will include a mix of Radeon Instinct graphics processing units and Epyc central processing units in a four-to-one ratio, meaning there will be four GPUs for every processor.

AMD isn’t supplying just any off-the-shelf chips for Frontier. Instead, the company said that both the CPUs and GPUs will be custom designs tailored for the DoE’s computing requirements. The chips are set to based on a different architecture than the company’s commercially-available products and will feature software likewise created specifically for Frontier.

The DOE’s decision to choose AMD over larger rival Intel Corp. for the project is something of a surprise. Patrick Moorhead, president and principal analyst at Moor Insights & Strategy, observed that “it has been years since AMD was inside the number one supercomputer, all the way back to the Opteron days,” referring to a processor family that the company launched in 2003.

The fact that the DOE will use AMD silicon to power what is likely to be the world’s fastest supercomputer is a valuable vote of confidence in the company. The news is especially significant given that Intel is involved in the U.S. government’s push for exascale computing, too.

“I believe the DOE chose AMD for three reasons: performance of the CPU and GPU piece-parts, the combined platform performance of CPU plus GPU enabled via Infinity Fabric [and] semi-custom silicon capabilities evident with the Microsoft Corp. and Sony Corp. engagements,” Moorhead added. Microsoft and Sony use customized AMD chips in their newest video game consoles.

The AMD-powered Frontier supercomputer will serve the DOE’s Oak Ridge National Laboratory, the operator of the current record-holding Summit system. The facility’s scientists plan to harness the platform for a variety of tasks ranging from medical research to manufacturing-related projects.

Oak Ridge staff have already started developing scientific applications for the system atop the existing Summit hardware. To aid their efforts, AMD and Cray are working on a a set of custom software tools for Frontier, including optimized versions of popular machine learning tools.

Photo: Randy Wong/Lawrence Livermore National Laboratory