CUDA minimal installation to save space

[Anderhar]

While using Sunshine to stream from a rented VPS, I'm concerned about saving rather pricey disk space, and full CUDA Toolkit seems like the biggest waste of it.

So here's my question: how about installing only the components specifically needed for Sunshine (some libraries?), without just running sudo apt install cuda-toolkit in order to look at all those unnecessary new icons in my application menu?

I've already tried to find some definitive list of Sunshine dependencies to experiment on my own, but all in vain so far.

[ReenigneArcher]

I'm not 100% sure on this, but I believe cuda is only a compile time dependency. If you (or any Linux user) could confirm this, that would be excellent.

[Anderhar]

Don't you mean that CUDA is needed only for those who avoid pre-made binaries and prefer to compile on their own? Because I used to think of CUDA as a system requirement for NVENC/NvFBC support. But I can actually test it this week on a planned clean re-installation.
(and thanks for the quick response)

[ReenigneArcher]

Yes, I think it's just a build time dependency, not a runtime one (but that might not be true).

That table indicates what cuda compatibility was used when we compiled the binaries we provide.

I don't have a Linux system with an NVIDIA card to test, so would be cool if you can confirm it!

[Anderhar]

Sorry for the delay.

Now I am happy to report that CUDA is really unnecessary for streaming on Linux in the NvFBC+NVENC mode: I just patched NVENC and voilà - everything is running! So your advice saved 8.4GB on my disk.

I think it's time to edit the docs by removing that whole misleading CUDA paragraph from "Getting Started" or correcting it in some "just for information" way.

The only thing that confuses me is a couple of errors in the Sunshine log, which, however, don't prevent streaming and may not even be related to the lack of CUDA:

Error: Couldn't release NvFBC context from current thread:  
Error: cuda::cuda_t doesn't support any format other than AV_PIX_FMT_NV12

[ReenigneArcher]

Thank you for confirming!

I don't think removing that is a good idea, as the version of cuda we compile with determines what GPUs are supported. But I agree that the wording can be improved so that users understand it is not required to install cuda themselves.

Regarding those two errors, I am not sure about that.