What is the current state of the CUDA implementation?

[Pafcifal95]

Some time ago I was looking into using the CUDA implementation of LightGBM instead of OpenCL, but according to the installation guide:

• "Only Linux is supported, other operating systems are not supported yet"
• "The CUDA implementation should be considered experimental, and we suggest using it only when it is impossible to use OpenCL version (for example, on IBM POWER microprocessors)"

These points are still in the installation guide but based on other issues posted here and the release notes I'm getting the impression that this may be outdated, so I just wanted to confirm:

• Is still only Linux supported, or is there also a CUDA build that works in a Windows environment?
• Is the CUDA implementation still considered experimental and should be avoided whenever possible, or can it be considered more mature/robust by now?

[jameslamb]

Thabks for your interest in LightGBM.

Yes the CUDA version is currently only supported on Linux.

But no, it's no longer "experimental" and you should prefer it to the OpenCL-based version if you have NVIDIA GPUs. Thanks for bringing that to our attention, we'll update those docs!

[Pafcifal95]

Thank you for the quick reply James! Any plans for a CUDA-based build for Windows?

[jameslamb]

No one is actively working on it, as far as I know. Would you like to help contribute that?