AMD Instinct GPU – 1.19 ExaFlops System and LLM

AMD has been designing high end Graphic Processors for a long period. In the 2010s, AMD joined the HPC project with DOE in the States and delivered the first ExaFlops system – Frontier. Frontier has not only achieved more than 1.1 Exa 64 bits Flops (2022 Jun), but also is the first system to deliver more than 60GFlops/Watt which provides extraordinary energy efficiency. With the breakthrough in performance, the comprehensive GPU software development environment named ROCm was used to provide necessary tools (Compiler, Debugger, Profiler) and software libraries for the software experts to develop GPU codes to run on AMD MI-Instinct GPU.

​AMD MI250X GPU was the GPU that used to build the Frontier system. The MI250X GPU Architecture has Hardware level Unify memory feature. CPU and GPU are connected with AMD XGMI Infinity fabrics. The following on generation MI300A, which is another breakthrough that integrated CPU and GPU in the same Socket to have all HBM shared between GPU and CPU which could eliminate the costly memory copy between System memory and HBM memory. Expected memory bound applications will be benefited by this architecture. Another coming product, MI300X, which follows OAM design standard, will provide 192GB HBM3 memory and 5.2TBs internal HBM memory bandwidth, that is suitable for LLM applications. AMD chip-let architecture has been proven as a successful technology that provides flexibility and scale up ​capability within a socket, and is able to deliver versatile processing/memory ​combination on a single Socket to support the growing demand of ​different workload​.