Originally Posted by Asisvenia
I agree with you, a 600mm Volta or Ampere gpu specifically dedicated to FP32 cores without Tensor cores and FP64 units would be same as Titan V or even much better.
I don't know is Nvidia going to change architectural plans but we need something for GeForce cards without those Tensor cores and FP64 units. For those who can't realize how much space had been taken by those units, check here:
It been said many times the TENSOR cores in the diagram is exaggerated and it doesn't take much as much die space as you would expect.
If you scale up P100, which doesnt have any TENSOR cores by 40%, you will get around the same die size as V100. Similarity if you take the transistor count of P100 and scale up by 40%, you will also get the expected transistor counts of V100.
Process node always strictly adhered to transistor per mm^2, and that itself show you TENSOR cores die space is literally non-existence.