Nvidia TensorCores, AMD Matrix Engines, Intel XMX Cores, all offer mat-mul acceleration in hardware, can be used for running CNNs as in Lc0, or in gaming for upsampling of video game resoultions. I think it is quite impressive to upsample 2K to 4K 3D w/o human-eye observeable difference, if you told me in ~2010 this will be psossible, I would say no way....
https://talkchess.com/forum3/viewtopic. ... 10#p910127
and, gamer and server brand GPUs differ meanwhile in architecture/feature set:
https://talkchess.com/forum3/viewtopic. ... 10#p911369
and, there are specific TPUs /ASICs out there, for optimized inference throughput:
https://en.wikipedia.org/wiki/Tensor_Processing_Unit
--
Srdja
