Moreover, not to mention the result is mostly valid for mod2 aritmetic; something the authors chose to mention it in the fine print causing this, a bit overblown in my opinion, excitement.
So for now it seems like we don't need to take action for regular matmul. On Wed, Oct 26, 2022, 16:03 Robert Kern <robert.k...@gmail.com> wrote: > On Wed, Oct 26, 2022 at 9:31 AM <canht...@gmail.com> wrote: > >> Hello! >> >> I was curious on how AlphaTensor will effect NumPy and other similar >> applications considering it has found a way to perform 3x3 matrix >> multiplication efficiently. >> https://www.deepmind.com/blog/discovering-novel-algorithms-with-alphatensor. >> I am not even sure how NumPy does this under the hood, is it 2x2? >> >> Is anyone working on implementing this 3x3 algorithm for NumPy? Is it too >> early, and if so why? Are there any concerns about this algorithm? >> > > numpy links against accelerated linear algebra libraries like OpenBLAS and > Intel MKL to provide the matrix multiplication. If they find that the > AlphaTensor results are better than the options they currently have, then > numpy will get them. In general, I think it is unlikely that they will be > used. Even the older state of the art that they compare with, like > Strassen's algorithm, are not often used in practice. Concerns like memory > movement and the ability to use instruction-level parallelism on each kind > of CPU tend to dominate over a marginal change in the number of > multiplication operations. The answers to this StackOverflow question give > some more information: > > > https://stackoverflow.com/questions/1303182/how-does-blas-get-such-extreme-performance > > -- > Robert Kern > _______________________________________________ > NumPy-Discussion mailing list -- numpy-discussion@python.org > To unsubscribe send an email to numpy-discussion-le...@python.org > https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ > Member address: ilhanpo...@gmail.com >
_______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
