Naive question: Most of ML workload in a datacenter runs on Linux, for which I assume there's good driver support. Why do we say that Nvidia does not have enough driver support on linux?
> Why do we say that Nvidia does not have enough driver support on linux?
It's a long story.
Nvidia has provided pretty-good datacenter/CUDA support on Linux for a long time, now. The problem is desktop Linux. Nvidia wanted to focus on supporting the old x11 desktop server at a time when most distros were switching to Wayland as a replacement. Nvidia tried to fix the issue by giving Linux devs a proprietary render API to develop Wayland support on, but it was largely rejected since it required writing a lot of platform-specific code. For a few years, the only way to properly accelerate Wayland on Nvidia hardware was to use the reverse-engineered Nouveau drivers that broke most desktop software - a catch-22 for Nvidia users that wanted a more updated desktop experience.
Very recently, a few things started changing. For one, Nvidia started to shuffle around their proprietary GPU code to make their hardware more open and modular. For two, starting with 510-series drivers and continuing through 535 and 555, Nvidia has started to make Wayland support a priority. In the long-term this should resolve the issue of Nvidia GPUs requiring special workarounds to support modern desktops.
They are pretty distinct. Having good CUDA drivers didn't mean raster graphics were accelerated well, and for a while the bulk of Nvidia's efforts were focused on a depreciated backend.
The good news is that those days are mostly behind us, with the more recent Nvidia drivers. Guess the crypto/AI boom helped them find the cash to hire better Linux devs.
It's a reasonable but naive question. There's a big difference between Nvidia offering driver/CUDA support for DC hardware vs providing driver support for desktop/gaming gear. There's also a big difference between the quality & consistency of driver support for desktop GPUs between Nvidia & AMD.
It's the very same driver. And one big selling point of NVIDIA has always been that you could run CUDA on your consumer card too - unlike AMD where their compute stack of the day would only run on a select few cards of the last generation.
It is mostly the same but the support that the DCs are getting are completely different. Nvidia can and does give them slightly tweaked drivers. Data centers ship a set of enterprise distros with tightly managed set of software that's been vetted. Since there are no binary guarantees in Linux, anything can happen with consumer hardware.
Source: once worked as an student system admin in my university HPC center.
As far as i know, Nvidia drivers are not stable enough. When a driver crashes in a data center the computer might reboot and continue the operation after some minutes. When a player plays a game and the graphic card crashes, he might readily switch to Windows.
Naive question: Most of ML workload in a datacenter runs on Linux, for which I assume there's good driver support. Why do we say that Nvidia does not have enough driver support on linux?