Hello,
Our hypervisors are based on KVM. I understand there is Tesla GRID support for KVM, but I have not see any reference implementations. Can you GPU-share on Tesla GRID + KVM by simply passing through the GPUs, or do you have to implement that yourself with the SDKs?
Any insights or documentation would be greatly appreciated.
Of course you can simply passthrough the GPU to guests, involves hiding the hypervisor, afaik.
The Nvidia supported way would be vGPU, currently no general OOTB solution, but RedHat is working on that, maybe ask them.
Can you pass-through individual GPU cores on a single card, or do you have to pass through the entire card?
Most likely, every GPU is its own PCI device. Use lspci to see them.
Right, but if I have 4096 CUDA cores split across 2 GPUs, I’m not going to to be able to pass-through individual cores… just GPUs, right? In other words, I won’t be able to give more than 2 VM’s GPU cores.
Right. If you want to do that, you need a vGPU setup.