4 ms·
CUDA is not the issue. AMD have already reimplemented like 80% of it, and honestly that part of it mostly works fine. Pytorch supports it, (almost) all the big
by FeepingCreature 2y ago
CUDA is not the issue. AMD have already reimplemented like 80% of it, and honestly that part of it mostly works fine. Pytorch supports it, (almost) all the big frameworks support it, if you're not doing really arcane things it just works. It's the drivers! They took like two years after the release of their flagship card to stop randomly crashing. Everything geohot has ever said about AMD drivers is 100% true. They just cannot stop shooting themselves in the foot.
- ryao 2y agoWhat did he say?
- FeepingCreature 2y agoGeohot (temporarily) giving up. https://github.com/ROCm/ROCm/issues/2198#issuecomment-1574383483 https://github.com/ROCm/ROCm/issues/2198#issuecomment-157438... Sadly most of the real spicy Twitter messages are gone since he deleted all his content, but there was a really fun one where he went off on a beautifully cryptic commit message in the driver. He also begged AMD to opensource the firmware so he could debug it. Sadly, AMD promised to do it and then nothing happened, as is typical for AMD promises. That's why tinygrad nowadays is aiming to just bypass the driver and firmware entirely.
- ryao 2y agoWho is tinygrad?
- FeepingCreature 2y agotinygrad = George Hotz. (It's his company, __tinygrad__ is basically his "work account".)