Why spmv use tensor core? #3
Description
Take fp64 for example, int A100 tensor core flops = 2x cuda core. However, AmgT just use 1/8 tensor core, so it's slower than cuda core.
{{ message }}
Take fp64 for example, int A100 tensor core flops = 2x cuda core. However, AmgT just use 1/8 tensor core, so it's slower than cuda core.