First Batch Of AMDGPU Changes For Linux 4.16: DC Multi-Display Sync, Vega Tuning

[polarathene]

@polarathene.
Sorry, my post was concerning the selling of gimped cards still marketed as full RX 560's. That's false advertising at the very least, and a below the belt move.
I know I wouldn't be very happy if I purchased a card and then found out it wasn't what I'd paid for but something a full tier below it!
That's why I'm going to hold off for a while until things become more clear.

Ah ok, kinda similar to 1060 I guess, except for the marketing it as a full 560 specs rather than 560 variant where you have to pay attention to specs listed on product page. I just tell people who ask about thoughts on nvidia 1060 to get a 1070, avoids them making a mistake if they're not into tech details too much and think askign for a 1060 at the store will be the same as any other 1060.

[pal666]

Did I miss AMD offering a full refund to those who bought RX560 GPUs

did you buy rx560 from amd?

[bridgman]

SR-IOV is a hardware feature iirc, I had a discussion about this on reddit with Bridgman months ago(or maybe it was on Phoronix), SR-IOV is similar to Nvidia's GRID or something.

SR-IOV is full HW virtualization on the GPU, with a separate set of on-chip state for each VF and the ability to reset as single VF independently of the rest. Don't believe any other vendors have this today.

I believe Intel's software implementation is described as "hardware assisted" in the sense that it makes full use of the IOMMU for isolation and efficient passthrough. Not sure about NVidia.

You'll more than likely find that AMD would have to implement some architecture change to properly support vGPU or bring SR-IOV to consumer cards.

AFAIK the issue is that the fusing we use to enable SR-IOV on PRO cards only supports on/off - I don't think we have the ability to pick something in-between (say 2-4 VFs rather than ~16 on a PRO card) for consumer.

[Brophen]

AFAIK the issue is that the fusing we use to enable SR-IOV on PRO cards only supports on/off - I don't think we have the ability to pick something in-between (say 2-4 VFs rather than ~16 on a PRO card) for consumer.

Maybe we could get a "ShaderRipper" Enthusiast Graphics card that has it "on"? :P Would love to have Threadripper + sr-iov.

[agd5f]

What is the use case you are imagining for SR-IOV? My impression is that you want to run your regular desktop on the physical function and various VMs on the virtual functions. That's not really possible. SR-IOV is designed for servers. The cards are generally headless and you can't really use the physical function as a regular device while the virtual functions are in use. You also can't use displays in the virtual functions. Only certain engines are virtualized.

[Brophen]

Lots of things! Some of them may be imaginary but, like Wendell says, being able to have apps from windows in container with 3d acceleration would be pretty nifty.

My imaginary use case would be to be able to have multiple different containers/vms, each accessing the same powerful graphics card, and being able to send each vm to a chosen display. That way you could have local multiplayer using one powerful machine. Especially helpful since few pc games support splitscreen, and why would you want splitscreen anyway if you can power multiple displays? But how many games support multi-display multiplayer? 2 might be a gracious guess.

or maybe you're a pentester that uses something like Kali linux that you don't use as a daily driver but still want gpu acceleration for certain tools

[polarathene]

What is the use case you are imagining for SR-IOV? My impression is that you want to run your regular desktop on the physical function and various VMs on the virtual functions. That's not really possible. SR-IOV is designed for servers. The cards are generally headless and you can't really use the physical function as a regular device while the virtual functions are in use. You also can't use displays in the virtual functions. Only certain engines are virtualized.

I'm interested for VM use with QEMU/KVM and VFIO. As Bridgman and yourself explained, probably not SR-IOV that is desired then, but whatever changes are needed to support the VFIO mediated device interface like Intel does? If that requires some hardware/architecture changes, then I guess it's too late for Navi to get that, let alone Vega? I don't just use VMs for gaming, so SR-IOV would still be useful for compute work I guess if I'm using a VM and a dev is using another, then if there is only one of us or we want to do an overnight compute task, can allocate the full GPU resources to one VM. No other uses come to mind atm, feel free to share other kinds, maybe they would apply to my workplace(lot of digital content creation and data processing).

[agd5f]

Lots of things! Some of them may be imaginary but, like Wendell says, being able to have apps from windows in container with 3d acceleration would be pretty nifty.

My imaginary use case would be to be able to have multiple different containers/vms, each accessing the same powerful graphics card, and being able to send each vm to a chosen display. That way you could have local multiplayer using one powerful machine. Especially helpful since few pc games support splitscreen, and why would you want splitscreen anyway if you can power multiple displays? But how many games support multi-display multiplayer? 2 might be a gracious guess.

or maybe you're a pentester that uses something like Kali linux that you don't use as a daily driver but still want gpu acceleration for certain tools

As per my comments about, the use cases you envision are not possible with our SR-IOV implementation. Much of what you want to do is possible today with bare metal.

[chithanh]

What is the use case you are imagining for SR-IOV? My impression is that you want to run your regular desktop on the physical function and various VMs on the virtual functions. That's not really possible. SR-IOV is designed for servers. The cards are generally headless and you can't really use the physical function as a regular device while the virtual functions are in use. You also can't use displays in the virtual functions. Only certain engines are virtualized.

Cross-plattform development of applications which use GPU acceleration, without having to reboot the system each time to test on the other platform. 2 VFs will be sufficient for this use case.
If that means I have to use the CPU integrated graphics for display, so be it.

[chithanh]

Buyers are not "confused", they are deliberately deceived.

The official specs at the time of purchase said that the card could come with 896 or 1024 shaders.

Sure, if you notice that you have been deceived. AMD and their accomplices now probably bet on "very few people following the IT press will notice they were deceived, so let's not proactively inform any buyer, and keep the swag!".

It has been reported on pretty much every tech website that covered the RX 560 launch. So where should the buyers have read the information about 1024 shader RX 560, that wouldn't tell them about the 896 shader model?