Intel To Use New Xe GPU Tech For First U.S. Exascale Supercomputer
Intel will use the company's new Xe GPU architecture to build the first exascale supercomputer in the United States, underlining the chipmaker's ambitions to up the competition with Nvidia in the high-performance computing space.
The Santa Clara, Calif.-based company said the supercomputer, named Aurora, is being developed for the U.S. Department of Energy at its Argonne National Laboratory in Chicago. The contract, which is slated for completion in 2021, is valued at more than $500 million.
The timeline provides a hint as to when Intel may release a data center product using its new Xe GPU architecture that was first detailed during the Intel Architecture Day event held in December. Intel has said that its first discrete graphics card product will release for consumers in 2020, and that products in other segments, including data center, will follow.
The Aurora supercomputer will be capable of performing 1 exaflop, or a quintillion floating point computations per second, Intel said. 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.
During a presentation, Rajeeb Hazra, the executive in charge of Intel's government and enterprise business, said that the upcoming Aurora supercomputer will use the company's new "Xe Compute Architecture," but declined to provide any details surrounding the technology itself.
He said the Xe Compute Architecture will "address new workload needs, particularly the kinds of workloads that will be coming in with the convergence of [high-performance computing, artificial intelligence] and data analytics."
Dominic Daninger, vice president of engineering at Nor-Tech, a Burnsville, Minn.-based high-performance computing systems builder, said he wouldn't be surprised if Intel avoided referring to a data center GPU product as a GPU because of its connotation with graphics cards.
Nvidia, for instance, refers to its Tesla processor for data centers as both a GPU and an accelerator.
"In many people’s minds, GPUs are just graphics," he said. "This is just a pure compute device here."
Intel said it will use the Xe Compute Architecture in combination with next-generation versions of the Intel Xeon Scalable Processor and Intel Optane DC Persistent Memory for the Aurora supercomputer. The company, which is working with supercomputing vendor Cray on the project, will also use its new One API software, which will provide developers to map software to specific hardware, such as a CPU, GPU or AI accelerator.
"This supercomputer will drive unprecedented innovation that accelerates the convergence of traditional [high-performance computing] data analytics and AI at exascale," Hazra, the Intel executive, said.