3 ms·
Care to explain how GPU virtualization is somehow worse/harder than CPU virtualization ?
by roblabla 9y ago
Care to explain how GPU virtualization is somehow worse/harder than CPU virtualization ?
- madez 9y agoI didn't say that. It's more oil to the fire.
- albertgoeswoof 9y agoBrb, turning off all my cpus and gpus, can’t be adding more fuel!
- madez 9y agoYou seem to misunderstand. It's not about not using CPUs or GPUs, but about what code runs on them, and how that code is accounted for. Who vetted for it and who is responsible for it?
- saltcured 9y agoConsidering that GPU architecture and low-level programming is still treated largely as trade secrets and the domain of proprietary blobs, it's hard to argue how virtualization should be trustworthy. When the OS designers get a good spec about the GPU instruction sets, memory architecture, MMU, and IO-MMU, and the OS designers can be the ones in charge of programming these to be consistent with existing OS protections, then we might start to approach an equivalence with CPUs. Then and only then, we can start worrying about subtle, modern risks like errata where the GPU hardware deviates from the idealized spec or where side-channel attacks take the wind out of our sails. Right now, we have something much closer to the GPU being another computer with an opaque and proprietary OS and a higher-level "managed code" runtime that accepts jobs defined in a mystery ISA by a proprietary compiler suite embedded inside the "graphics driver" in our own computer. We have to have blind faith that there is any data protection as well as that there aren't any code-injection flaws that would allow malicious data and "shader" code to hijack the GPU and its bus interface to compromise the host computer.