Hacker News Reader

Comment by freedomben | Hacker News Reader

It seem self-defeating that they don't. If nvidia fully opened up their driver, they would de-facto own the GPU market (at least for ML, but probably gaming as well). Conversely if AMD could get support into most of the ML libs, they could start cleaning up.

View on HN

> Conversely if AMD could get support into most of the ML libs, they could start cleaning up.

This is happening, albeit slowly, fwiw. Pytorch has ROCm support, but not great. AMD did win the exascale machine contracts and have custom hardware in the machines that are built. It does look like they've been using those contracts to invest more into computational sciences/ML. I can only hope that they continue at the pace they have in the consumer market. Competition between AMD and Nvidia (and hopefully Intel -- a pipe dream) can only be beneficial to us in the ML community.

They already own the data center gpu market, what are you taking about?

Last I checked they already had monopoly evoking shares of that market. Like 80+ or 90+ percent

There's a better chance of hell freezing over, but it would certainly benefit users. From a economic standpoint though, it's entirely unsurprising why they do this. CUDA owns, and in the AI age it's hard to beat the parallelism Nvidia offers. If Nvidia cedes their acceleration advantage, they're leaving themselves open for attack. DirectX is a funny example of this, still being closed-source despite a reverse-engineered version existing with often better performance.

It would be nice if everyone writing UNIX-like drivers gave away their source code, but then the books wouldn't balance. It's a tale that's half a century old now.

Reply on news.ycombinator.com