At the ISC 2021 Supercomputing Conference and Exhibition, AMD revealed updates to its open ROCm platform, unveiled its Education and Research (AIER) support initiative for using AMD Instinct gas pedals, and demonstrated the adoption dynamics of its EPYC processors in the High Performance Computing (HPC) industry.
The latest Top500 list of supercomputer rankings shows continued growth in AMD EPYC processors for HPC systems. AMD EPYC processors now power nearly five times as many systems (49 vs. 10) compared to the June 2020 list. In addition, they are used in half of the 58 new systems added to the current ranking. High-performance computing is becoming increasingly important in solving many of the world's most important problems. And with our EPYC and Instinct products, AMD is committed to delivering the kind of performance and capabilities that will break the exascale barrier, accelerate scientific discovery and innovation;& ; commented Forrest Norrod, senior vice president, Data Center and Embedded Systems Division, AMD. The latest HPC systems built on AMD products include CSD3, COSMA8, Discoverer, Perlmutter, MeluXina, the UK weather bureau hybrid system, four cloud NDv4 clusters in Microsoft Azure, Singapore's National Supercomputer Center (NSCC) computing complex, the US National Center for Atmospheric Research (NCAR) system and others. However, about half of new systems with EPYC in the latest TOP500 list are paired with NVIDIA gas pedals, which usually provide the main performance in such combination.
A 2020 Intersect360 survey of HPC users found that AMD EPYCs left a positive impression on 78% of respondents. By comparison, in 2016 there were only 36%. According to a new Intersect360 survey, 23% of respondents said they use EPYC processors extensively in HPC systems, and another 47% are testing or using these processors to some degree. The recent launch of AMD's new EPYC 7003 series of processors is likely to contribute even more to this trend. The company did not show any new gas pedals at ISC 2021, but to help scientists and researchers take advantage of all the power of the AMD Instinct solutions already in the portfolio, the AIER educational program is called for. It offers remote access to Instinct, AMD ROCm training center and corresponding software, as well as technical support and manuals on AMD software and hardware solutions. Both regional partners and global companies participate in the program, including Dell Technologies, Gigabyte, HPE and Supermicro. Open software platform ROCm continues to receive industry support, acquiring new applications, libraries and frameworks to use AMD gas pedals. This demonstrates that AMD's proposed CUDA to C++ conversion tool, HIP, has been embraced by the community as a heterogeneous programming model that can be used when writing or adapting their codes to accelerate on AMD GPUs, including Gromacs, TensorFlow and GridTools. In particular, PyTorch for ROCm is now available as a Python installable package, opening up opportunities for developers to do mixed-precision computing and large-scale learning using AMD's MIOpen and RCCL libraries. And more recently, CuPy, an open source library for GPU computing, received version 9.0 with support for the ROCm stack and AMD gas pedals.