Implementation - radumarias/rvirt-gpu GitHub Wiki

Virtual GPU API: using wgpu create an abstraction layer for GPU instructions
vGPU (GPU driver): create a GPU driver that virtualizes the GPU, as a vGPU
- bindgen_cuda
- rusticl
- nova
- mps
- VirtIO/GL
- NVIDIA vGPU
- mesa
- QEMU
- duplicating the llvmpipe gallium driver, Mesa's Gallium architecture
Local app: decrypts the data with the key from OS's keyring and runs on vGPU
Network Communication Layer: Using tokio-rustls send the instructions to GPU-coordinator over TLS
GPU-Coordinator: distribute the instructions to multiple real GPUs and sends back the results over TLS.
Server-Side GPU Execution
- using ash
- Rust-CUDA to run the instructions on the actual GPUs. GPU-Coordinator sends back the results over TLS.
- we can monitor the GPU with nvml-wrapper
Handling Results: On the client side, deserialize the received results and make them available to the local application through the Virtual GPU API.