Nvidia are doing something like this now advertising their CUDA TOPS as double their actual value but with a footnote telling you this is for "sparse" multiplic

Nvidia are doing something like this now advertising their CUDA TOPS as double their actual value but with a footnote telling you this is for "sparse" multiplications. Here "sparse" means each dot-product can assume half of each input vector is zero.