AMD charges back into the HPC fray with new systems

After years of watching its presence shrink on the Top 500 supercomputer list, AMD is battling back with a new set of EPYC-based server processors and specially-tuned GPUs for high-performance computing (HPC) in a complete server system.

The company and its partners announced new servers with the EPYC 7601 processor, which it claims is three times more performance-efficient than Intel’s best Xeon server processors, the Xeon Platinum 8180M1, as measured by SPECfp[i] benchmark. The news came at the Supercomputing ’17 show taking place in Denver.

Target workloads for AMD solutions include machine learning, weather modeling, computational fluid dynamics, simulation and crash analysis in aviation and automotive manufacturing, and oil and gas exploration, according to the company.

OEMs offering AMD’s EPYC systems

The company has an impressive lineup of OEMs and system integrators offering EPYC systems: Asus, BOXX, Gigabyte, HPE, Inventec, Penguin, Silicon Mechanics, Supermicro, Synnex, and Tyan. Notably absent: Dell EMC and Lenovo. However, Dell did make a positive statement of support for EPYC when it was launched, so Dell may be working on something and it just wasn’t ready to announce yet.

Also announced was ROCm 1.7., the latest version of the open-source GPU HPC development platform. It now includes multi-GPU support for the latest Radeon GPU hardware, as well as support for TensorFlow and Caffe in the MIOpen libraries.

AMD partners are showing off or announcing systems at SC17. ASUS, Gigabyte Technology, Penguin Computing, Supermicro and TYAN all announced servers based on EPYC, while Inventec showed off its P47 system, which it launched in August. The P47 combines an EPYC CPU with four Radeon Instinct MI25 GPUs, each delivering up to 12.3 TFLOPS of single precision performance.

AMD

AMAX showed off the [SMART]Rack P47, an appliance featuring 20 P47 racks to provide up to a PetaFLOP of single precision compute performance and more than 10 terabytes of DDR4 memory per rack. It should be noted that a lot of HPC loads are double precision, which is much slower than single precision.

The [SMART]Rack P47 also features HPC-optimized [SMART]DC DCiM software for remotely monitoring, managing, and orchestrating GPUs, especially in situations where temperature, power and system health are particularly important to operation.

AMD has done a good job building support for EPYC. Now, it’s up to the partners to sell it. As I said in a prior blog post, it takes a while for these things to make their way into the market — at least a year. That’s especially true with HPC, since it takes so long to qualify them. But so far, so good.