According to a recent leak, AMD is working on the Instinct MI250X, a next-generation graphics accelerator that will use the CDNA2 architecture, which means that its GPU will have an MCM design, that is to say, of type multichip module. This means that we will no longer see a monolithic core GPU, but will combine at least two chips to create a “super GPU”.
The first information we are seeing points to a significant jump in gross performance. The AMD Instinct MI250X will use a 7nm in-process GPU, for a total of 110 computing units, in addition to 128GB HBM2e memory. These two keys already allow us to intuit that we are facing a solution that will be capable of offering a very high performance, but we are going to delve a little deeper into this issue.
Those 110 computing units would translate into a total of 7,040 shaders under the current model of 64 shaders per CU (Unit of Computing), but that figure doesn’t quite match the massive performance we see on both FP32 and FP64. I think this indicates that AMD could surprise us with a major change at the structure level, and that the Sunnyvale company could double the number of shaders per CU, going to a total of 128.
It’s unconfirmed, but it’s the only thing that makes that raw power data make sense. And speaking of raw power, as we see in the attached image, the AMD Instinct MI250X would reach 383 TFLOPs in FP16 (medium precision), 47.9 TFLOPs in single precision and 47.9 TFLOPs in double precision.
As for the HBME memory, this model will come equipped, as we said, with 128 GB, a whopping four times the maximum of the AMD Instinct MI100, which adds up to 32GB of HBM2. We do not have details on the bandwidth, and the design of the cooling system is not known, but it is likely that AMD will take advantage of the jump to the next generation to introduce important changes.
Otherwise, the AMD Instinct MI250X will have a TDP of 500 watts and GPU working frequencies of 1.7 GHz. We don’t have a definitive release date yet, but we expect it to compete with upcoming Intel and NVIDIA solutions (Ponte Vecchio (Xe-HPC) and NVIDIA H100) as well. which is likely not to arrive until mid-2022.