Tensor Core utilization in cuDSS

Accelerated Computing GPU-Accelerated Libraries
,
1

Hi,

I would like to know if tensor core is utilized in cuDSS. Since cuBLAS uses tensor core for dense matrix operations and cuDSS is based on cuBLAS, I assume cuDSS also uses tensor core.

I tried to profile both float and double matrix solves. However, I did not see tensor core usage in sm80_xmma_gemm_f32/sm90_xmma_gemm_f64 functions. Does this mean cuDSS does not support tensor core at the current stage?

Thank you.
Di

2

Hi Di!

We consider this to be an implementation detail. We don’t disable usage of tensor cores and we might use them (depending on the matrix structure and the induced internal data structure), either through cuBLAS or with our own kernels.

Whether tensor cores are used or not, depends on the matrix.

Let us know if you want to know more details but I am not sure how those would help a user of cuDSS.

Thanks,
Kirill

1 Like