Formula for Fermi Core Count

On the Tesla architectures I multiply “Multiprocessor Count” by 8 to get the “Core Count”, what is the formula for it on a Fermi



From my understanding it executes two warps concurrently with 16 SPs on each warp. That is 2 cc’s for each warp so 32 threads are being executed each cc. Somebody correct me here.

Yes, multiply by 8 for compute capability 1.x and 32 for compute capability 2.0.

And for compute capability 3.0 it will be a factor of PI to the power of 4.

Hmm, really I find this disappointing that there is no future proof way of finding out the total number of CUDA cores based on the deviceQuery API. If the core count keeps changing with each major compute capability revision, nVidia should expose either the total core count or a core multiplier (per SM) in the deviceQuery. Only then it will be possible to estimate the device’s peak computational power.