Client GPU support (both APUs and dGPUs)

[DemiMarie]

This would be incredibly useful for many purposes, including gaming.

[Quedale]

I can't say how much I second that.

Allowing decent GPU virtualization could finally be the final building block to fully support a multi-desktop personal gaming computer.
With the possibility of restoring split-screen gaming to its former glory.

Looking forward to find out what comes next!

[DemiMarie]

@xxmtg which ones are those?

[saadrahim]

Only some special GPUs has hardware SRIOV capability. We are expanding the support to as many as possible that available on the market!

@xxmtg Please identify yourself as it appears unlikely that you are an AMD employee. You are not a member of any AMD GitHub organization. We may remove your post as it appears you may be misleading people.

[corigne]

Stumbled on this while doing my periodic check into SRIOV support on commodity GPUs. Unofficial word appears to be commodity dgpus are "in the pipeline". Is that the case? If so, how far down the road are we talking, and how can we help?

[DemiMarie]

Intel will be offering SR-IOV in its Panther Lake iGPUs and in its Arc Pro dGPUs, so this could help avoid losing customers to Intel. That said, I expect that upstreaming the drivers (#2) would be critical for that, as many of these potential customers will not want to use out of tree drivers.

[amd-mwevers]

Hi DemiMarie,

Our MxGPU software requires SR-IOV functionality to be enabled in silicon, which to date has only been included and enabled on GPUs sold to cloud and datacenter channel customers. While there is some overlap between datacenter and consumer silicon, we are not planning to introduce fragmented support by enabling MxGPU on consumer devices that could theoretically support SR-IOV.

We appreciate you raising this topic as it helps us keep track of customer demand and can influence product planning; however, at this point, I don't have any updates on this topic to share.

[corigne]

Hi DemiMarie,

Our MxGPU software requires SR-IOV functionality to be enabled in silicon, which to date has only been included and enabled on GPUs sold to cloud and datacenter channel customers. While there is some overlap between datacenter and consumer silicon, we are not planning to introduce fragmented support by enabling MxGPU on consumer devices that could theoretically support SR-IOV.

We appreciate you raising this topic as it helps us keep track of customer demand and can influence product planning; however, at this point, I don't have any updates on this topic to share.

I'm not sure if this is the result of poor media reporting or some change in internal planning, but didn't someone recently mention discrete GPU support was roadmapped for these drivers? I know I read about this not that long ago here: https://www.phoronix.com/news/AMD-GIM-Open-Source

This is something the commodity GPU community has been very interested in for a while now, as is indicated by increased support from other vendors in this space (Intel and Nvidia both). Most alternatives to multi-guest mediated graphics device sharing are clunky or outright prohibitive. Almost all of them require having a second GPU or an APU onboard and others require the use of hacky scripts to manage host/guest transitions. It's really not an ideal situation at the moment, and with increasing linux desktop market share official support for this would be desirable.

What can we, as the community, do to help bring this about?

[DemiMarie]

If you are using KVM as hypervisor and Linux as the guest OS, virtio-GPU native contexts allow allow one to use AMD, Apple M1/M2, and Qualcomm GPUs with near-native performance for graphics workloads. To the best of my knowledge, ROCm, non-KVM hypervisors, and non-Linux guests are not supported. ROCm will be able to be supported eventually, but the compute-only XNACK feature will not be supported due to Linux kernel memory management limitations.

[Quedale]

If you are using KVM as hypervisor and Linux as the guest OS, virtio-GPU native contexts allow allow one to use AMD, Apple M1/M2, and Qualcomm GPUs with near-native performance for graphics workloads. To the best of my knowledge, ROCm, non-KVM hypervisors, and non-Linux guests are not supported. ROCm will be able to be supported eventually, but the compute-only XNACK feature will not be supported due to Linux kernel memory management limitations.

Work is currently underway. (see here)
It's far from stable, but I got d3d9 games to launch with very good performance!

I think DRM Native Context might be the best consumer accessible solution I've seen so far.

[amd-mwevers]

I'm not sure if this is the result of poor media reporting or some change in internal planning, but didn't someone recently mention discrete GPU support was roadmapped for these drivers? I know I read about this not that long ago here: https://www.phoronix.com/news/AMD-GIM-Open-Source

He was referring to the Radeon PRO V710 which is our datacenter / cloud GPU. MxGPU-Virtualization project launched with Instinct GPU support only (Radeon PRO V710 was delayed for stability fixes). To be clear - there is currently no plan to add RDNA3 / RDNA4 support in this project, except for the Radeon PRO V710.