AMD said this week that it had started initial shipments of its next-generation Instinct compute GPUs based on the CDNA2 architecture. These aren’t the sort of GPU you’d find in the best graphics cards. Instead, the new Instinct MI200 compute accelerator carries 128GB of HBM2E memory and is designed for supercomputers and other high-performance computing applications, such as AI or ML.
“Initial shipments of next-generation AMD Instinct accelerators featuring 2nd Gen CDNA architecture,” an AMD statement in an earnings presentation reads.
AMD’s Instinct MI200 ‘Aldebaran’ compute GPU will power the world’s first exascale-class Frontier supercomputer, due to be delivered in late 2021 (as well as some smaller machines), so it’s about time for AMD to start production of the accelerator. AMD has not yet disclosed any architectural peculiarities about CDNA2, so the only two things we know about the Instinct MI200 for sure is that it features two dies connected with each other using AMD’s Infinity Fabric, and it carries 128GB of HBM2E memory that connects to the dies using two 4096-bit interfaces.
AMD’s codenamed Instinct MI200 ‘Aldebaran’ compute GPU is expected to more than double performance of the Instinct MI100 (184.6 FP16 TFLOPS, 23.1 FP32 TFLOPS, 11.5 FP64 TFLOPS) and probably add support for new data formats as well as instructions. Yet, we do not know how the CDNA 2 architecture differs from the original CDNA — which actually derives from AMD’s GCN-based Vega 7nm that was optimized for compute.
Some leakers indicate that the MI200 could have as many as 16,384 stream processors across two of its chiplets, which is more than double the MI100’s 7680 active stream processors. Given such a significant ‘brute force’ improvement, and assuming that there are tangible architectural enhancements, it is more than logical to expect the MI200 to be more than two times faster than its predecessor. Yet, since this information is unofficial, take it with precaution.
AMD did not disclose whether it began initial shipments of its Instinct MI200 compute GPU to the Oak Ridge National Laboratory, the owner of Frontier, or some other customer(s). Nonetheless, the initial shipments indicate that the accelerator is good enough to be shipped to clients.