The code lists GFX906_60, GFX908_120, and GFX90A_110, which should be the product line of AMD Instinct series computing cards. Among them, GFX906_60 and GFX908_120 should correspond to Instinct MI60 and Instinct MI100. The number at the end indicates the number of CUs provided by the GPU. Previously, it was rumored that the internal code name of the CDNA 2 GPU
architecture was GFX90A, so GFX90A_110 in the code corresponds to the chip of the CDNA 2 architecture, providing 110 CUs. Since the new generation of products will use MCM (Multi-Chip-Module) packaging technology, the Instinct MI200 computing card will have two “computing units”, and the complete solution will provide 220 CUs.
The number of CUs indicated in these codes does not represent the complete configuration of the GPU. In order to have a better yield rate and ensure the supply quantity, AMD will disable some CUs. It is understood that the number of CUs for each “computing unit” of the Instinct MI200 computing card should be 128. Assuming that the Instinct MI200 computing card is configured with 220 CUs and the core frequency is 1500 MHz, then the single-precision computing performance is 42.2 TFLOPS, which is 1.82 times that of the current Instinct MI100 computing card.
Rumor has it that the Instinct MI200 computing card adds support for full-rate FP64 computing capabilities, the corresponding performance maybe two to four times that of the Instinct MI100 computing card, which means that there will be better performance in some HPC workloads
In addition to MCM packaging technology, the new generation of products will also adopt the next generation of Infinity Fabric bus technology, and support the third generation of AMD Infinity architecture, allowing up to 8 channels of the same GPU to be connected to Exascale. The Instinct MI200 computing card will be used in the new generation of supercomputers in the United States, such as Frontier at Oak Ridge National Laboratory and El Capitan at Lawrence Livermore National Laboratory.