AMD Instinct MI250X computing card FP64 performance has been greatly improved, 7nm process is still used

As early as May this year at the 49th Global Technology, Media and Communications Conference held by JP Morgan Chase, AMD CEO Dr. Lisa Su has confirmed that the GPU with the CDNA 2 architecture codenamed Aldebaran will be launched within this year. AMD’s Frontier, supercomputers built for the US Department of Energy’s Oak Ridge National Laboratory (ORNL), will use this Instinct computing card.
Recently, Twitter user @ExecuFix revealed that the new GPU based on the CDNA 2 architecture will be manufactured using a 7nm process, and the rumored Instinct MI200 series of computing cards will have two products, namely Instinct MI250 and Instinct MI250X. Among them, the Instinct MI250X will be equipped with 110 CUs and 128GB of HBM2E video memory, with a frequency of 1.7 GHz and a TDP of 500W. Since this GPU will use MCM packaging technology, I am not sure if 110 CUs refer to a single computing module or the sum of two computing modules. The new generation of products will also use the next-generation Infinity Fabric bus technology and support the third-generation AMD Infinity architecture, allowing up to 8 channels of the same GPU to be connected to Exascale.

In addition, some data is provided. The single-precision (FP32)/double-precision (FP64) calculation performance of the Instinct MI250X calculation card is 47.9 TFLOPs, and the half-precision (FP16/BF16) calculation performance is 383 TFLOPs. The FP64 computing performance is 4.16 times that of the Instinct MI100 computing card. As a reference, the FP64, FP32, and FP16 computing performance of the NVIDIA A100 80GB version are 19.5 TFLOPs, 156 TFLOPs, and 312 TFLOPs respectively.