That’s the one thing that I feel is a bit misleading in the article (to be fair, it was initially written years ago, and got rewritten a bit recently). FLOPS co

That’s the one thing that I feel is a bit misleading in the article (to be fair, it was initially written years ago, and got rewritten a bit recently). FLOPS comparisons given in the wild are not always apple-to-apple (eg. not including Tensor cores for NVIDIA, but including V_DUAL_DOT2ACC_F32_F16 for AMD), while on the flip side, AMD’s WMMA should address the same goals as Tensor cores. I have an article on comparing the two: https://espadrine.github.io/blog/posts/recomputing-gpu-performance.html https://espadrine.github.io/blog/posts/recomputing-gpu-perfo...