Setting up a small(ish) CUDA runtime environment in Debian

Accelerated Computing CUDA CUDA Setup and Installation
1

My current goal is to run gpu-burn from a bootable USB stick.

What works right now is to follow the instructions at the CUDA Toolkit 12.8 Downloads site, and install cuda-toolkit-12-8 and nvidia-open. gpu-burn compiles and runs fine, identifies the GPUs correctly - all good.

The problem is that apt-get -y --no-install-recommends install cuda-toolkit-12-8 nvidia-driver-cuda nvidia-open installs over 10GB of files - which is a lot for a USB stick.

Is there a minimal set of packages that I could install without pulling in things that I don’t need, like graphical utilities, profilers, debuggers, etc.?

Thanks.

2

Is there a reason you need to compile gpu-burn from the USB stick?

Why not compile it, enabling all GPU architectures you’re interested in and then just copy the binary to the USB stick.

3

Good point - I already have the binary from previous runs, and could just copy it to the USB stick.

All I would need is the drivers and any libraries that the executable depends on. Is there a set of Debian packages that does only that? Even the nvidia-open package by itself wants to install Nsight.

4

I’m not familiar with gpu-burn or the packaging system, but I’d have thought all you need is a driver to suit the range of cards. That should contain all that’s required.

5

To some extent, gpu-burn is a read herring. As long as I can run any CUDA app, I’m fine.

The problem is the packaging system; I guess I’ll have to hunt down all the necessary libraries one by one if I want to keep things small.

6

You could just download the driver directly from nvidia.com. It will come with all required runtime bits.

7

Looks like that’s the way to go. The cuda-drivers package plus the cuBLAS library seem to be sufficient.

Thanks.