Re:Nvidia Unveils GeForce RTX 30 Series GPUs
by nl on Tuesday September 01, 2020 @11:47PM
(#24349805)
> Giving away that secret sauce in their drivers, so that AMD could make their cards seamlessly compatible would probably be a huge mistake from a business perspective for Nvidia.
I've seen this argument a lot but never from the ML perspective - and I don't think it makes sense with ML.
Nivida's advantage is ML is CUDA and cuDNN, and the huge set of tools built on them. These aren't the driver, but a layer above it.
I don't really understand the "secret sauce" argument for the drivers at all, but I assumed it applied to gaming somehow. I can't think of how it applies to ML though.
Re:Nvidia Unveils GeForce RTX 30 Series GPUs
by pjmlp on Wednesday September 02, 2020 @01:35AM
(#24350275)
PTX is on the driver level, and it took NVidia several years effort to make their cards properly map into C++ memory module.
Intel and AMD despite their FOSS support, actually only do open what they feel like.
Re:Nvidia Unveils GeForce RTX 30 Series GPUs
by fluffything on Wednesday September 02, 2020 @04:15AM
(#24351118)
nvidia drivers work just fine on wayland with kde and gnome, both of which support EGL and GBM.
The other key benefit would be driver management. Updating your drivers right now can be a nightmare. If the drivers are open sourced and upstreamed into the kernel this becomes a non-issue. I do understand why Nvidia doesn't want to do that though. A lot of their lead right now is not just hardware based, but software based. They have a choke-hold on the ML ecosystem and it's a huge cash cow for them. Giving away that secret sauce in their drivers, so that AMD could make their cards seamlessly compatible would probably be a huge mistake from a business perspective for Nvidia.