SCUDA: CUDA-over-IP

SCUDA is a CUDA over IP bridge allowing GPUs on remote machines to be attached to CPU-only machines.

Setup

Using prebuilt binaries / images

You can find prebuilt binaries in the release tab There is also a Docker image ghcr.io/k0in/scuda:latest which can be used. For an example setup you can look at docker-compose.yml.

Build locally

Build using earthly

earthly +all

if you want to cross compile (supported: linux/arm64 and linux/amd64) the binaries, you can also run

earthly +cross

If you want to build it using plain docker you can also build the provided docker file.

docker build --build-arg BASE_IMAGE=nvidia/cuda:12.6.2-devel-ubuntu24.04 . -t scuda:local

Usage

To use SCUDA, ensure the server is running on a machine with a GPU:

scuda-server

or using docker

docker run --gpus all -p 14833:14833 ghcr.io/k0in/scuda:latest

The server port can be changed using the environment variable SCUDA_PORT.

On the client machine, set the SCUDA_SERVER environment variable to the server's IP and set the SCUDA_PORT if modified, and preload the SCUDA library:

export SCUDA_SERVER=<ip>:port
LD_PRELOAD="libscuda.so" nvidia-smi

Acknowledgements

I (K0IN) did nothing exepct writing a dockerfile, readme and a build that i can understand.

ALL CREDITS GO TO THE AWESOME WORK OF kevmo314 the original author of this project. thank you for your work.