Matrix multiply breakthrough, AlphaTensor (could also do for other algorithms): "AlphaTensor discovered algorithms that are more efficient than the state of the art for many matrix sizes."

[stevengj]

@stevengj

even 20% speedups are not worth the price of allocating the auxiliary matrices required by Strassen-like algorithms (especially since the BLAS interface doesn’t provide any way for the user to pass in auxiliary buffers to re-use)

I should note that there was a 2016 paper that argued otherwise, claiming that optimized matrix-multiplication libraries already use such buffers internally—at least in the context of libflame and BLIS—and that the same buffers can be used for Strassen.

So, it may be that there are some practical BLIS-based linear-algebra libraries that currently use Strassen for sufficiently large sizes, e.g. ffpack and tblis, and these may be fruitful to update to use the new AlphaTensor-discovered algorithms.