The ncu report is from gemm test on a A100.
I want to know what the “Op/s” means.
And I notice that the peak value of “Tensor Core Operation Roofline” is four times that of “Floating Point Operations Roofline (Half Precision)”.Does “Op” means 1/4 half flop?
Op here means each multiplication and addition as part of the tensor core mma instruction. A100 has a peak fp16/bf16 dense tensor core throughput of 312 Tflops ( NVIDIA A100 | NVIDIA ), i.e. 3.12E+14 to match the roofline. This corresponds to 1024 multiply-adds per SM per cycle (or 256 per Tensor Core).
Thank you!!!