Re:Matrix multiplication inches closer to mythic goal
by dekhn on Tuesday March 23, 2021 @01:34PM
(#26557716)
in the time spent improving matrix multiplication by a tiny incremental amount, many gigawatt hours have been spent making existing CPUs, GPUs, and TPUs do matrix multiplication using far less efficient algorithms. It's unclear this work would really be important in production computing (I understand the theoretical desire) compared to time spent making other things work better.
[ Parent
| on HN ]
Re:Matrix multiplication inches closer to mythic goal
by MauranKilom on Tuesday March 23, 2021 @04:30PM
(#26560021)
> It's unclear this work would really be important in production computing
It most certainly won't [0]. But neither will something like 99% of mathematical research right now.
[0]: See first example in https://en.wikipedia.org/wiki/Galactic_algorithm
Re:Matrix multiplication inches closer to mythic goal
by snapcore on Tuesday March 23, 2021 @03:03PM
(#26558922)
I was under the impression the Strassen method was rarely used because it has a number of drawbacks such as harder to parallelize, uses more memory, has bigger constants for complexity, less stable, etc.
It's not really designed to be optimal for the current hardware, but just to lower the number of steps. It could be more useful in the future though.