19 Apr AMD And Cray Join Forces To Bring EPYC To HPC
Cray and AMD have announced the availability of AMD EPYC 32-core CPUs on the supercomputer company’s midrange cluster system, the Cray CS500. AMD has enjoyed a long and successful relationship with Cray, supplying Opteron processors in big iron Cray supercomputers such as that deployed in the National Center for Supercomputing Application (NCSA) Blue Waters at the University of Illinois at Urbana-Champaign, and the massive $97M 18,688 CPU Titan system at the Department of Energy’s Oak Ridge National Laboratory. This new announcement is hardly surprising given this shared history, but it does mark an important milestone for AMD as it seeks to regain some of its lost luster in the HPC market.
What did AMD and Cray announce?
In a brief announcement, the companies said that the AMD EPYC processor is now supported on the CS500 with four dual-socket nodes per 2U chassis. AMD also announced it is adding a single two-socket model with a large memory capacity that should benefit many HPC workloads. AMD supports up to four memory channels per socket, compared to Intel’s support for three. As AMD seeks to find markets for EPYC, I’ve often thought that it should focus on HPC, where early adopters can adapt and optimize their codes fairly quickly to take advantage of EPYC’s unique floating-point performance and large memory subsystem. Notably, Cray has also ported the Cray Programming Environment (tools and libraries) to EPYC, which is a great addition to the AMD EPYC story in HPC. The new systems will start shipping this summer.
Unfortunately, from my perspective, the new AMD-quipped CS500 missed the opportunity to support AMD GPUs. Cray supports NVIDIA GPUs in the CS Storm line for HPC and AI applications. I would suspect that the AMD Vega GPUs could make an attractive addition to the lineup if a large customer sought out its support. Perhaps this will happen when AMD launches their 7nm GPU later this year or next.
Figure 1: The Cray CS500 Cluster Supercomputer enables Cray to sell medium-size systems based on industry-standard x86 technologies
While this announcement in itself is hardly a game changer, AMD EPYC’s prospects in HPC could become a significant factor in its battle with Intel Xeon chips in the market. While many industry observers anxiously await ARM-based chips from Cavium and Qualcomm to cut into Intel’s dominant market share, I for one believe that AMD EPYC, with its Intel Xeon instruction set compatibility, stands a good chance of gaining share—especially if the company is able to execute on a successful and timely 7nm second-generation EPYC SOC. If AMD is able to leverage its GPUs and the critical software ecosystem for HPC and Machine Learning, called ROCm, to provide a more balanced and optimized platform for performance-hungry applications, it stands a very good chance.