AMD and Cray have announced that they’re building “Frontier,” a new supercomputer for the Department of Energy at Oak Ridge National Laboratory. The goal is to deliver a system that can perform 1.5 exaflops: 1.5×10¹⁸ floating point operations per second.

By way of comparison, a single Nvidia RTX 2080 GPU manages about 14 teraflops of compute performance with 32-bit numbers. Frontier will achieve 100,000 times more. The fastest supercomputer in the Top 500 list weighs in at 200 petaflops, or 0.2 exaflops. As things stand, it’d take the top 160 machines on the list to match Frontier’s performance.

Frontier will use custom versions of AMD’s Epyc processors (likely Zen 3 or Zen 4), matched with 4 GPUs, all connected using AMD’s Infinity Fabric. Between nodes, Cray’s Slingshot interconnect will be used, which has transfer rates of up to 200Gb/s per port. The GPUs will have their own stacked HBM (High Bandwidth Memory). It’ll be housed in 100 cabinets, taking about 7,300 square feet of floor space. Power consumption will be 30-40MW.

The plan is for Frontier to be delivered in 2021, at a cost of about $500 million for the hardware and $100 million for research and development. It should be the fastest supercomputer in the world when it’s delivered, and it will be the US government’s second exaflops-capable system; the first will be the 1 exaflop Aurora, built using Intel Xeon SP processors and Intel Xe GPUs.

The supercomputer will be made available to academics to run a wide range of simulations and experiments. The Radeon Instinct GPUs include hardware dedicated to machine-learning workloads, and it’s likely that Frontier will be used for this kind of task in addition to the more conventional weather and nuclear weapon simulations.

https://arstechnica.com/?p=1501655