https://devblogs.nvidia.com/parallelforall/inside-pascal/
Compute Capability 6.0 Table.
Kepler, Maxwell, and Pascal GP100 have exactly the same number of resources available to particular amount of cores.
With Kepler its 192 cores, with Maxwell its 128, with Pascal its 64 cores. In essence 64 cores in Pascal are doing the same job, as 192 cores in Kepler, because they are not starved for resources. This is the reason also, why GTX 980 was faster than GTX 780 Ti in compute. Similar story we see with Similarly core counted Titan X, and P40. P40 is faster, despite having the same amount of cores, and lower core clock.
What about consumer Pascal? Each SM in GP102, 104, 106 GPUs have 128 cores. And they are fed, by the same amount of resources, as previous versions of Nvidia architecture. Exactly the same as Maxwell, but with higher core clocks. This is what I meant when I have written that Pascal and Maxwell consumer architectures are no different. If you will OC GTX 980 Ti to GTX 1080 levels, the difference in performance is very ver close.
How come? GTX 980 Ti, 2816 CC GPU, clocked at 1492 MHz, and 2090 MHz on Memory:
136 FPS.
GTX 1080 - reference design, with 320 GB/s and 2560 CC's
137 FPS. With higher core clocks.
This is the reality of Consumer GPUs. I am dumbfounded that you never knew this was the case. Only new Pascal GPU was GP100 chip. GP102, and the rest of consumer GPUs were just Maxwell GPUs on 16 nm node.
P.S. None of the consumer GPUs, nor professional GPUs made with GP102, 104, 106 chips have the new features from GP100, chip(49-bit memory, unified memory and FP16). Guess why?
