Known Issues

• Running an agent inside a Docker container requires host networking, which is not supported in some environments (i.e. Docker Desktop, macOS, and VM-based containers). WSL sharing GPUs has the same limitations as Docker WSL and is subnetted with the host machine.

• DX11/DX12/OpenGL/Vulkan are the only supported graphics APIs. Most graphical software should work out of the box. CUDA 10-12.4 is supported for compute. PyTorch and TensorFlow binaries also work out of the box.

• Resolutions beyond 4K may have questionable performance due to video encoder/decoder limitations. This will be addressed with a software encoder in the future.

• Graphics applications rendering significantly above 60 FPS (i.e. benchmarks like FurMark) will be limited by video encoder capabilities rather than GPU rendering. This may result in degraded performance at very high frame rates compared to native execution.

• Applications that are being run with juice will not fall back to a local GPU on the client when the remote GPU disconnects.

• Using a log level that's not ERROR on the CLI may severely degrade performance.

• Client applications that leverage a remote GPU for juice may take a few minutes to close when the GPU is disconnected, depending on the application. This can cause confusion if another GPU is connected in the meantime.

• Running the juice login command from a Linux shell may trigger an output from the browser launching (e.x. "Failed to load module 'canberra-gtk-module'", etc.). You can safely disregard this output.