AMD Launches Arcturus As The Instinct MI100, Radeon ROCm 4.0

AMD is marking the SC20 virtual conference this week by launching the AMD Instinct MI100 accelerator, which is based on their CDNA architecture. Also notable and coinciding with the MI100 launch is the Radeon Open eCosystem 4.0 (ROCm 4.0) Linux release.

The AMD Instinct MI100 accelerator is based on their CDNA architecture and what we've come to know as Arcturus over the past year of their Linux driver support coming together. AMD believes the MI100 is the world's fastest HPC GPU accelerator and can surpass 10 TFLOPS of FP64 compute power. The MI100 is rated for 11.5 TFLOPS of peak FP64 performance and up to 46.1 TFLOPS of FP32 matrix performance. The MI100 paired with AMD Matrix Core technology is rated for nearly 7x boost to its FP16 peak performance for AI training compared to previous generation AMD accelerators.

The AMD Instinct MI100 makes use of 32GB HBM2 memory at a 1.2GHz clock rate and capable of 1.23TB/s memory bandwidth. MI100 supports PCI Express 4.0 connectivity and packs 120 compute units and 7680 Stream processors. CDNA is an evolution of the Vega architecture rather than RDNA/Navi that is gaming optimized rather than compute. Given the HPC focus and how long we've been seeing the Arcturus Linux patches mature, the Linux support for the AMD Instinct MI100 support should be in great shape for launch albeit we haven't been able to test the accelerator to confirm its Linux support state.

On the ROCm 4.0 front, this updated Radeon Open eCosystem stack is designed to make it easier than ever to port NVIDIA CUDA code-bases to it and doing so in a performant manner. ROCm 4.0 is being advertised as an open-source platform for the "Exascale Era" and supports both the OpenMP 5.0 industry standard and their HIP interfaces along with the likes of the PyTorch and Tensorflow frameworks. One of the focuses moving forward for ROCm is making it easier for developers to deploy and port their software to running on this open-source compute stack rather than the proprietary confines of NVIDIA CUDA. In regards to ROCm 4.0, will be firing it up soon on other GPUs in looking at its performance.

AMD is also using SC20 to say that EPYC "Zen 3" processors are shipping in volume to select HPC and cloud customers this quarter. The public launch of AMD EPYC Zen 3 CPUs will be in Q1'2021.