I’m using Quadro RTX 6000 ( Turing)

I want to know using tensor core or not.

If the kernel name is the format below such as “volta_fp16_s884gemm”, is it right to use the tensor core?

s[some digits]

If so,

Am I using Tensor Core?

I’m using Quadro RTX 6000 ( Turing)

I want to know using tensor core or not.

If the kernel name is the format below such as “volta_fp16_s884gemm”, is it right to use the tensor core?

s[some digits]

If so,

Am I using Tensor Core?

yes, that is a kernel that uses tensorcore

In addition you can use a profiler to confirm this. On the “old” profilers, the metric is `tensor_precision_fu_utilization`

. On the “new” profilers (i.e. nsight compute), the corresponding metric is `sm__pipe_tensor_op_hmma_cycles_active.avg.pct_of_peak_sustained_active`

then volta_scudnn_128x64 , implicit_convolve_sgemm also use tensor core?

And

I can’t find the value(sm__pipe_tensor_op_hmma_cycles_active.avg.pct_of_peak_sustained_active) on the result of profiling (Figure below)

What should I do?

Learn to use the profilers to retrieve metrics. Any of the linked blogs shows how to do that. There are literally hundreds or thousands of metrics. They are not all shown in the summary report you have excerpted.

I don’t know. It’s not obvious to me that it does. Kernels with 884 or 1664 in the name usually indicate that way that they use tensor core. I’m not sure about every kernel, however. But you can use the profiler method to be sure.