3 ms·
> The sad part is that, by all appearances, the goal of this work was not to add functionality for NVIDIA GPUs in particular. I don't see how that can be true
by IshKebab 6y ago
> The sad part is that, by all appearances, the goal of this work was not to add functionality for NVIDIA GPUs in particular.
I don't see how that can be true given that it is designed to speed up ML, and only nVidia graphics cards are used for ML. Approximately nobody uses Intel or AMD.
Maybe they meant it wasn't written by an nVidia employee? But it's clearly intended to be used with nVidia GPUs in particular.
- deft 6y agoIt uses an NVIDIA apis, its made for NVIDIA tasks (CUDA) and it was only ever designed to work on NVIDIA cards. Apparently being a Facebook employee means you can't have pro-NVIDIA anti-anything else intentions... even when that's clearly the case.
- mroche 6y agoApparently being a Facebook employee means you can't have pro-NVIDIA anti-anything else intentions... even when that's clearly the case. That’s being a tad disingenuous. Developing a solution to an in-house use case while ignoring alternative platforms that you don’t use does not make one “anti-anything else”. For a business of Facebook’s scale (or any scale for that matter), would you or your manager approve of spending time on such a task? Unless it was simple, there was enough time, or there were known potentials of switching product lines, I would venture to guess: no. However, trying to upstream such work without considering where it’s going in which the very last patch review of 21 shows the problem isn’t the brightest idea.