search

edera-dev / krata

krata is an isolation engine for securing compute workloads
Apache License 2.0
220 stars 10 forks source link

feature: GPU acceleration via virtio-GPU native contexts #191

Open DemiMarie opened 1 week ago

DemiMarie commented 1 week ago

This allows for GPU acceleration without the firmware/hardware nightmare that is PCI passthrough. It is in production in ChromeOS for Qualcomm GPUs, under development for Intel, AMD, and GPUs, and neither requires nor uses GPU hardware virtualization support. It works by forwarding the GPU kernel driver API, which is already a security boundary. Most kernel driver vulnerabilities are mitigated, as ioctl submission is serialized and arguments are at least somewat validated.

Qubes OS will be using native contexts for GPU virtualization (https://github.com/QubesOS/qubes-issues/issues/8552) and I am interested in collaborating on the implementation. I presented on this at Xen Project Summit 2024.

azenla commented 1 week ago

This could be very interesting! Our use case is slightly different in that we need to support things like CUDA, would that be tenable in this situation? If not, we may have to still retain support for NVIDIA vGPUs in our commercial product.

DemiMarie commented 6 days ago

This could be very interesting! Our use case is slightly different in that we need to support things like CUDA, would that be tenable in this situation? If not, we may have to still retain support for NVIDIA vGPUs in our commercial product.

I’m not very familiar with how Nvidia vGPU works, but my understanding is that it uses mediated devices. Does the vGPU Manager use a kernel module (other than nvidia.ko)?

azenla commented 6 hours ago

It uses the standard kernel modules, so nvidia.ko and friends :)