Cray and AMD are coming back to Oak Ridge National Laboratory to partner on the world’s largest and most expensive supercomputer. The Department of Energy’s Oak Ridge National Laboratory has selected American HPC company Cray–and its technology partner AMD–to provide the lab with its first exascale supercomputer for 2021 deployment.

The $600 million award marks the first system announcement to come out of the second CORAL (Collaboration of Oak Ridge, Argonne and Livermore) procurement process (CORAL-2). Poised to deliver “greater than 1.5 exaflops of HPC and AI processing performance,” Frontier (ORNL-5) will be based on Cray’s new Shasta architecture and Slingshot interconnect and will feature future-generation AMD Epyc CPUs and Radeon Instinct GPUs.

In a media briefing ahead of today’s announcement at Oak Ridge, the partners revealed that Frontier will span more than 100 Shasta supercomputer cabinets, each supporting 300 kilowatts of computing. Single-socket nodes will consist of one CPU and four GPUs, connected by AMD’s custom high bandwidth, low latency coherent Infinity fabric.

Oak Ridge Director Thomas Zacharia indicated that 40 MW of power, the maximum power draw set out in the CORAL-2 RFP, would be available for Frontier.

“Cray’s Slingshot system interconnect ties together this massive supercomputer and a new system software stack fuses the best of high performance computing and cloud capabilities,” said Cray CEO Pete Ungaro. “We worked together with AMD to design a new high density heterogeneous computing blade for Shasta and new programming environment for this new CPU-GPU node.”

Frontier will use a custom AMD Epyc processor based on a future generation of AMD’s Zen cores (beyond Rome and Milan). “[The future-gen Epycs] will have additional instructions in the microarchitecture as well as in the architecture itself for both optimization of AI as well as supercomputing workloads,” said AMD CEO Lisa Su, adding that the new Radeon Instinct GPU incorporates “extensive optimization for the AI and the computing performance, [with] mixed-precision operations for optimum deep learning performance, and high bandwidth memory for the best latency.”

The CPU and GPUs will be linked by AMD’s new coherent Infinity fabric and each GPU will be able to talk directly to the Slingshot network, enabling each node “to get the optimum performance for both supercomputing as well as AI,” said Su. All these components were designed for Frontier but will be available to enterprise applications after the system debuts, according to AMD.

Frontier marks a return for Cray and AMD to Oak Ridge, home to another Cray-AMD system, Titan. Benchmarked at 17.6 Linpack petaflops, Titan was the number one system in the world when it debuted (as an upgrade to Jaguar) in 2012. With Titan set to be decommissioned on August 1, 2019, and Frontier scheduled to be deployed in the back half of 2021 and accepted in 2022, Oak Ridge won’t be without a Cray-AMD machine for too long. While Titan used AMD (Opteron) CPUs and Nvidia (K20X) GPUS, Frontier will rely on AMD for all its in-node processing elements.

Frontier is Oak Ridge’s third machine to use a heterogeneous design. In addition to the aforementioned Titan, Oak Ridge is of course home to Summit, which became the world’s fastest supercomputer in June 2018. Its 143.5 GPU-accelerated Linpack petaflops are owed to 9,216 Power9 22-core CPUs and 27,648 Nvidia Tesla V100 GPUs.

“Since Titan, Oak Ridge has pioneered this idea of having GPU accelerators along with CPUs,” said Zacharia. “Frontier will be the third generation of supercomputing system built around this architecture and it will be the second generation AI machine.”

Frontier will be used for future application simulations for quantum computers, nuclear energy systems, fusion reactors, and precision medicines, said Zacharia, adding “Frontier finally gets us to the point where we can actually design new materials.”

In terms of number-one system bragging rights, the DOE has previously stated, and recently confirmed, that Aurora (aka Aurora21, the revised CORAL-1 system that Intel is contracted to deliver to Argonne) is on track to be the United States’, and possibly the world’s, first exascale system in 2021; and since that messaging has not changed, we believe it is the intention of the DOE to deliver on that goal. However, even if it is the case that Intel keeps to its timeline and Aurora is deployed and benchmarked first, Frontier is slated to be stood up on a very similar timeline and according to publicly stated performance goals will provide roughly 50 percent more flops capability.

Asked to comment on the “competitive” timelines for Frontier and Aurora, Zacharia said he could only comment on Frontier.

“I don’t know all the details of Aurora procurement because that information has not been publicly released, but we do know that Frontier will be the largest system by far that the DOE has procured,” he said.

“We know that Oak Ridge has experience with Summit and Titan previously in using CPU-GPU systems. We also know that the pre-exascale system that the scientific community is using today to develop all their applications and system software is on our system Summit, which is the largest machine available to anybody…. If there is any competition between the labs, it’s just competition for ideas, which is what scientists should do, but otherwise this is truly a DOE lab system effort to ensure the United States maintains the forefront of this important technology, not only because it drives technology innovation in the IT computing space but it also drives economic competition and creates jobs.”

Zacharia further cited that the goals for Frontier are aligned and consistent with the White House AI initiative as well as the National Council on American Workers, which is creating new jobs using AI and scientific computing in manufacturing and other spaces.

As for that $600-million-plus price tag, it is “by far the most expensive single machine that [the DOE has] ever procured,” said Zacharia. It’s also Cray’s largest contract ever.

The total amount includes the system build contract for “over $500 million,” as well as the development contract for “over $100 million” that will, according to Ungaro, be used to develop some of the core technologies for the machine, as well as a new programming environment that will enhance GPU programmability via extensions for Radeon Open Compute Platform (ROCm).

“The Cray Programming Environment (Cray PE)…will see a number of enhancements for increased functionality and scale,” said Cray. “This will start with Cray working with AMD to enhance these tools for optimized GPU scaling with extensions for Radeon Open Compute Platform (ROCm). These software enhancements will leverage low-level integrations of AMD ROCmRDMA technology with Cray Slingshot to enable direct communication between the Slingshot NIC to read and write data directly to GPU memory for higher application performance.”

To support the converged use of analytics, AI, and HPC at extreme scale, “Cray PE will be integrated with a full machine learning software stack with support for the most popular tools and frameworks.”

Frontier marks Cray’s third major contract award for the Shasta architecture and Slingshot interconnect. Previous awards were for the National Energy Research Scientific Computing Center’s NERSC-9 pre-exascale Perlmutter system (with partners AMD and Nvidia) and the Argonne National Laboratory’s Aurora exascale system (with Intel as the prime).

Frontier is the first CORAL-2 award, announced nearly 13 months after the RFP was released. As laid out in the program’s RFP, CORAL-2 seeks to fund up to three exascale-class systems: Frontier at Oak Ridge, El Capitan at Livermore and a potential third system at Argonne if the lab chooses to make an award under the RFP and if funding is available. Like the original CORAL program, which kicked off in 2012, CORAL-2 has a mandate to field architecturally diverse machines in a way that manages risk during a period of rapid technological evolution. The stipulation indicates that “the systems residing at or planned to reside at ORNL and ANL must be diverse from one another,” however the program allows Oak Ridge and Livermore labs to employ the same architecture if they choose to do so, as in the case of Summit and Sierra, which employ very similar IBM-Nvidia architectures.

The CORAL-2 effort is part of the U.S. Exascale Computing Initiative. The ECI has two components: one is the hardware delivery and the other is application readiness. The latter is the domain of the Exascale Computing Project (see HPCwire‘s recent coverage to read about the latest progress), which is investing $1.7 billion to ensure there’s an exascale-ready software ecosystem to get the most from exascale hardware when it arrives.