ROCK-Kernel-Driver VS vgpu_unlock

No, just dangling documentation on the ROCm/HIP etc repos/sites, and the knowledge that all device DMA used to only support PCIe-a-like within the drivers/firmware, but as per that whitepaper above mentions a lot of effort was put into infinity fabric of CDNA2 so it would be very strange to use it for p2p. That the driver memory topology only recently enabled p2pDMA from what I have been following. This of course would require firmware (if not already there) and userspace to handle such a situation. Hence why I was asking for clearer source on MI250x specifically, and if it was recent.

Even if people refer to AMDGPU driver as AMD made, it is not AMD but guys from open source, this: https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver which also leads to this: https://www.kernel.org/doc/html/latest/ and https://en.wikipedia.org/wiki/AMDgpu_(Linux_kernel_module)#:~:text=Community%5Bedit%5D,libdrm%2C%20Xorg%2C%20Wayland#:~:text=Community%5Bedit%5D,libdrm%2C%20Xorg%2C%20Wayland). If i got that right. AMDGPU PRO driver is what is installed on Windows a proprietary driver which it seems is built upon AMDGPU open source but this one is signed by AMD... Witch got me thinking as AMD refusing to bring Radeon Software Center to Linux, seems like best thin for open source is to community and companies like Canonical, Rad Hat etc start builtin hardware for open software... a perfect mix of things?

Plus, if we're going to talk about stuff developers are sick of, how about releasing packages that are broken? https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver/pull/117 Kind of pointless to put up a release package if it's going to be obviously broken to this extent. In the end getting this to work still involved manually going in and patching things, both with the driver and then with the runtime libraries for tensorflow, making it not much better than dealing with NVIDIA. Even worse since this at least would've been caught in automated testing.

https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver/blob/roc-1.9.x/drivers/gpu/drm/amd/amdgpu/amdgpu_pm.c#L469

This isn’t even the first time a hacked driver has been used to unlock some HW feature - https://github.com/DualCoder/vgpu_unlock

Alternatively enable vGPU for the 2070 and use it for both Jellyfin LXC and Windows VM. https://github.com/DualCoder/vgpu_unlock

I'm on Linux and I'm running a 3070 Ti (Nvidia). I have always wanted to do GPU virtualization but because NVIDIA won't release vGPU for consumer card no one can do it without crossing legal red tape or problems with bricking your GPU. I did find this [https://github.com/jamesstringerparsec/Easy-GPU-PV] however it is only for windows, I found this [https://github.com/Arc-Compute/LibVF.IO/] and does not work with my GPU, and this [https://github.com/DualCoder/vgpu_unlock] and can't get it to work. Done any one know an alternative on Linux that work just like this, overcoming these problems (on KVM)?

Otherwise, your only other option is the real hardware virtualization options that are available. NVIDIA's enterprise vGPU solution is for expensive compute cards however some have had good luck making vGPUs work on consumer NVIDIA cards with tools such as vgpu_unlock

There was a hack to enable it on some consumer cards, but it’s not available on Ampere/30x0 cards: https://github.com/DualCoder/vgpu_unlock/issues/8

Pci passthrough https://github.com/mbilker/vgpu_unlock-rs https://github.com/DualCoder/vgpu_unlock https://github.com/DualCoder/vgpu_unlock/issues/91 https://gitlab.com/polloloco/vgpu-proxmox https://github.com/joeknock90/Single-GPU-Passthrough https://gitlab.com/YuriAlek/vfio#start-here https://www.reddit.com/r/homelab/comments/b5xpua/the_ultimate_beginners_guide_to_gpu_passthrough/ https://forum.level1techs.com/t/single-gpu-passthrough-with-proxmox/113282/2 https://forum.proxmox.com/threads/problem-with-gpu-passthrough.55918/