This is amazing. It sounds like the summary is that they turned matrix multiplication of various sized matrices into a kind of game, then had AlphaZero, the game playing AI, play the game and find optimal strategies. The new optimal strategies for “winning” the game in less moves were actually new matrix multiplication algorithms, some of which had been discovered previously, some of which were new. Many thousands of algorithms were found across a range of matrix sizes.
And of course aside from the intellectual curiosity, our GPUs are mega matrix multiplication machines and these new algorithms could be tuned to specific hardware and improve the overall graphics and ML performance by 10-20% on existing hardware.
And of course aside from the intellectual curiosity, our GPUs are mega matrix multiplication machines and these new algorithms could be tuned to specific hardware and improve the overall graphics and ML performance by 10-20% on existing hardware.
Cool!