It looks like the AMD Instinct MI200 graphics accelerator will receive 110 CUs. This is indicated by a commit in the machine learning library code on GitHub.
The new accelerator is believed to be hidden behind the identifier GFX90A_110. And for Instinct IM60 with identifier GFX906_60 and for Instinct MI100 with identifier GFX908_120, the last number in the identifier means the number of CUs in the GPU.
Thus, the Instinct MI200 is likely to receive 110 computing units. Like the Arcturus GPU that powers the MI100, the Aldebaran GPU for the MI200 will probably have 128 CUs, just not all of them will be involved. Let us remind you that Aldebaran will be a dual-GPU GPU.
It is impossible to speak about the number of stream processors unequivocally since we do not know the configuration of the new GPU. If it does not change as a whole, then we will talk about 8192 stream processors, of which 7040 will be active. If AMD does the same trick in the CDNA2 architecture as Nvidia does in the case of Ampere, then we can talk about 14 170 stream processors.
For a specialized accelerator, which will be the Instinct MI200, a very important parameter is the performance on double-precision operations (FP64). And here, according to rumors, Instinct MI200 will be much faster than Instinct MI100, since the performance in FP64 mode will be identical to that for FP32. This means that, depending on the architecture, the Instinct MI200 will be twice or even four times more productive than the Instinct MI100.
0 Comments