AMD Leverages EPYC And Vega To Cram 1 PetaFLOP Supercomputer Into A Single Server Rack

Back in the mid-2000s, AMD and IBM collaborated on a supercomputer project for the Los Alamos National Laboratory in New Mexico, which featured over 64,000 dual-core Opteron processors and numerous custom accelerators to achieve 1 PetaFLOP of compute performance. For a time, the massive system – dubbed Roadrunner -- was the fastest supercomputer on the planet. It required 696 racks and covered approximately 6,000 square feet.

Last night, at its Capsaicin event in Los Angeles, adjacent to SIGGRAPH which is currently underway, AMD unveiled a supercomputer with similar performance featuring its recently released EPYC processors and upcoming Vega GPUs, but this time the system was small enough to be rolled out on stage and filled only a single rack.

The Project 47 supercomputer was powered by 20 AMD EPYC 7601 processors and 80 Radeon Instinct GPUs. Other hardware included 10TB of Samsung memory and 20 Mellanox 100G cards (and 1 switch). All told, Project 47 is capable of 1 PetaFLOP of single-precision compute performance or 2 PetaFLOPS of half-precision.

Project 47 is built around the Inventec P47. The P47 is a 2U parallel computing platform designed for graphics virtualization and machine intelligence applications. A single rack of Inventec P47 systems is all that was necessary to achieve 1 PetaFLOP, and it does so while producing 30 GigaFLOPS/W, which AMD claims is 25% more efficient than some other competing supercomputing platforms

Thanks to its 32-core / 64-thread EPYC processors and Radeon Vega GPUs, which feature 4,096 stream processors each, AMD also claims that Project 47 rack has more cores/threads, compute units, I/O lanes and memory channels in use simultaneously than in any other similarly configured system.

The P47 is expected to be available from Inventec and their principal distributor AMAX sometime in Q4 of this year.