[David Kanter on Tech Report] - Nvidia VR preemption "possibly catastrophic".

[VR Enthusiast]

https://www.youtube.com/watch?v=tTVeZlwn9W8&feature=youtu.be&t=1h21m35s

David Kanter talks to Scott Wasson on the TR Podcast and confirms Nvidia's problems with Maxwell for VR.

He says that in terms of preemption they are "possibly catastrophic" and even behind pre-Skylake Intel. D:

 

[jpiniero]

All the more reason to buy Pascal, amirite?

 

[RussianSensation]

But it doesn't matter in practice because according to some users who only buy NV from one generation to the next, future proofing for VR, DX12 on Kepler or Maxwell is irrelevant since by the time these features are actually widely utilized/needed, they will have upgraded to Pascal or Volta, and so on. There is certainly some truth to this because realistically speaking next generation 2016-2018 DX12 games should in theory bring more advanced graphical effects/features that will overwhelm existing GPUs. Having said that, a lot more gamers are now prolonging the time between GPU upgrades, and thus I am not sure if this argument is valid for the general PC gaming population. What makes it more disingenuous in my eyes is that NV was marketing full DX12 support and having the most future-proof DX12 GPUs but now we find out that Maxwell's AC engines are broken/possibly catastrophic for DX12 and VR? Not cool as this is GTX970 marketing fiasco #2:

Had NV not done that, all of this wouldn't be such a big deal but it seems the current hardware may or may not back up the claims of their marketing team/slides.

Seeing a November 2013 $699 780Ti being outperformed by a $280 GTX970 in September 2015 is a eye-opener; and as always a great reminder why buying $600+ flagship cards in hopes of "future-proofing" for 4-5 years is a waste of $. This has been true for as long as I can remember.

That's why I am generally against recommending $600-1000+ GPUs for 'future-proofing' beyond 2, maybe 2.5 years. If a gamer cannot afford to buy flagship cards every single generation, it's far better to buy a $300-350 GPU and in 2-3 years another $300-350 GPU rather than buying a $600-700 flagship card and keeping it for 4-5 years. In that sense getting an R9 290/290X/390/970 as a stop-gap GPU (unless you can find a B-stock GTX980 for $370) seems like the smarter bet right now. I think that goes for both AMD and NV because Fury X is limited to 4GB HBM, while Maxwell's performance in VR/DX12 is uncertain. Next generation should bring 8GB HBM2 (and higher), and most likely far more capable DX12/VR architectures from both AMD and NV and a more advanced video decoding/encoding engine and possibly other features.

 

[ViRGE]

He says that in terms of preemption they are "possibly catastrophic" and even behind pre-Skylake Intel. D:

Much respect for David Kanter, but this seems hyperbolic, or at least taken out of context?

Intel doesn't support async compute (concurrent graphics and compute contexts) even on Skylake, let alone before it. Kepler couldn't be worse than Skylake, let alone Maxwell 2.

(And if we want to get really technical, NVIDIA has high priority contexts, HintCreateLowLatencyDevice, specifically to handle Timewarp on their earlier GPUs.)

 

[tential]

Rs to me vr doesn't matter at all right now. Because I'll have a new gpu by then in Arctic islands or pascal. No matter the gpu you have a lot of people may be in that boat. Even dx12 isn't that big of a deal for me (obviously it's a huge deal....) since Im sure nvidia will make their cards work in games wit dx12 support and if I played those new games, I'd want Arctic islands or pascal.

The users who are downplaying async I Don get though. This is good for gpu development moving forward and provides more immersive lighting and physics... Without it we won't advance as fast as humanly possible. I am happy for dx12/async because it will improve games. Which vendor who does it better next gen isn't relevant to me. I don't care who wins because I'll win. Because I just buy my best option.

Oh, closed on a $200 r9 290 trix locally. Wish I had known guskline was doing a flat ship but still happy to have a great sapphire card. Two in a row lol kind of funny. We'll see if sapphire gets my money for Arctic islands too.

 

[3DVagabond]

I understand people who say DX12 performance doesn't matter at this point for them because they will buy a new card before it becomes relevant. What I don't understand is why the".1" in DX12.1 was so important when they thought it meant that nVidia supported DX12 better than AMD.

People also need to understand that not everyone, probably not even a majority, upgrade every gen. and quit trying to dismiss it as "it doesn't matter".

 

[Black96ws6]

That's why I am generally against recommending $600-1000+ GPUs for 'future-proofing' beyond 2, maybe 2.5 years. If a gamer cannot afford to buy flagship cards every single generation, it's far better to buy a $300-350 GPU and in 2-3 years another $300-350 GPU rather than buying a $600-700 flagship card and keeping it for 4-5 years. In that sense getting an R9 290/290X/390/970 as a stop-gap GPU (unless you can find a B-stock GTX980 for $370) seems like the smarter bet right now. I think that goes for both AMD and NV because Fury X is limited to 4GB HBM, while Maxwell's performance in VR/DX12 is uncertain. Next generation should bring 8GB HBM2 (and higher), and most likely far more capable DX12/VR architectures from both AMD and NV and a more advanced video decoding/encoding engine and possibly other features.

Well said. Any of these cards: 290x/390/390x/970/980/980ti will hold you for a while. $300-$350 seems to be the sweet spot.

And some brick and mortar stores will price match online stores, so you don't have to wait, or worry about something happening in the mail, and if something goes wrong with it, you can take it right back to the store instead of the whole ship it back or RMA process.

And while it now looks like the whole AC on DX12 Nvidia thing was overblown, even if it wasn't, what DX12 games are you not capable of playing RIGHT NOW. There aren't any!

By the time there's a decent amount, you'll have upgraded anyway lol.

I am playing older games these days anyway. Civ5, CS, NS2, Def Grid, etc, with some Ark and Alien Isolation sprinkled in.

I had a 290x and it worked great for all those. Then it broke (luckily within the 30 day window), I took it back and got a GTX 970 SSC and FOR MY PARTICULAR games (fine print), it works even better. But we all know Nvidia is better at DX11 so no surprise.

Point is, these future proof discussions are really getting out of hand. Anyone that buys or bought a 290+/970+ is set for a bit. Period.*

* At 1080 and 1440 resolutions of course

 

[VR Enthusiast]

Much respect for David Kanter, but this seems hyperbolic, or at least taken out of context?

Intel doesn't support async compute (concurrent graphics and compute contexts) even on Skylake, let alone before it. Kepler couldn't be worse than Skylake, let alone Maxwell 2.

(And if we want to get really technical, NVIDIA has high priority contexts, HintCreateLowLatencyDevice, specifically to handle Timewarp on their earlier GPUs.)

I have no idea, I don't follow Intel GPUs much at all definitely not on the VR front. I didn't even know they were talking to Oculus but that's what he says.

I think Skylake can preempt on compute threads. I wonder if there could there be any other reason why he would say what he said?

 

[Silverforce11]

Much respect for David Kanter, but this seems hyperbolic, or at least taken out of context?

Intel doesn't support async compute (concurrent graphics and compute contexts) even on Skylake, let alone before it. Kepler couldn't be worse than Skylake, let alone Maxwell 2.

(And if we want to get really technical, NVIDIA has high priority contexts, HintCreateLowLatencyDevice, specifically to handle Timewarp on their earlier GPUs.)

They do have "priority context", they mentioned it in their GameWorks VR PDF. But there's a problem.

It's not a priority context. You know why?

Cos it has to WAIT for graphics in front of it to finish rendering. That by definition, cannot be priority context as the purpose of that usage is to get the Async Timewarp through traffic in the pipeline to lower latency.

p31. NV is very specific to developers on how best to minimize their problem.
https://developer.nvidia.com/sites/...works/vr/GameWorks_VR_2015_Final_handouts.pdf

And now we're getting insights that the queues are software scheduled, async is software scheduled via drivers, it is a disaster when it comes to low latency environments like VR.

 

[Arachnotronic]

By the time this is relevant, I'd bet that NVIDIA will have sorted things out.

That said, it really is interesting just how "future proof" AMD's GCN seems to be.

 

[monstercameron]

Much respect for David Kanter, but this seems hyperbolic, or at least taken out of context?

Intel doesn't support async compute (concurrent graphics and compute contexts) even on Skylake, let alone before it. Kepler couldn't be worse than Skylake, let alone Maxwell 2.

(And if we want to get really technical, NVIDIA has high priority contexts, HintCreateLowLatencyDevice, specifically to handle Timewarp on their earlier GPUs.)

If i'm not mistaken that data was from a source at occulus.

 

[Vesku]

I don't think it's hyperbolic since he was specifically referencing the experiences of Oculus in implementing VR. If Pascal still has that same limitation then they will have a major VR limitation and to a lesser extent DX 12 limitation until at least Volta. So it just depends on how quickly VR becomes important to high end GPU purchasers.

 

[n0x1ous]

Oxide just updated everyone on their conversations with Nvidia on Async. Apparently the driver doesnt have it fully turned on yet.......

"We actually just chatted with Nvidia about Async Compute, indeed the driver hasn’t fully implemented it yet, but it appeared like it was. We are working closely with them as they fully implement Async Compute. We’ll keep everyone posted as we learn more.

Read more: http://wccftech.com/nvidia-async-compute-directx-12-oxide-games/#ixzz3kpBL8yVQ

 

[tviceman]

Techreport is obviously ANTI-AMD, as has been stated and confirmed by the many AMD hardcore fans on the forums here, so for them to be reporting Nvidia-anything being "potentially catastrophic" actually means Nvidia is going out of business tomorrow. Good game Nvidia, nice knowing ya.

 

[iSkylaker]

VR lol
http://store.steampowered.com/hwsurvey/videocard/

people ain't got time for that.

 

[AnandThenMan]

Oxide just updated everyone on their conversations with Nvidia on Async. Apparently the driver doesnt have it fully turned on yet.......

The question is why. Maybe with it fully enabled performance will tank, or maybe Nvidia for some mysterious reason doesn't have full driver support. Which would be odd considering Nvidia has said they have been working on and contributing to DX12 for years. Either way the argument that DX12 doesn't matter at all on current hardware is nonsense.

 

[VR Enthusiast]

Techreport is obviously ANTI-AMD, as has been stated and confirmed by the many AMD hardcore fans on the forums here, so for them to be reporting Nvidia-anything being "potentially catastrophic" actually means Nvidia is going out of business tomorrow. Good game Nvidia, nice knowing ya.

Alternative: It was a live podcast so Wasson wasn't reporting anything. With his nervous laughter he looked like he wanted to cry.

 

[Silverforce11]

Alternative: It was a live podcast so Wasson wasn't reporting anything. With his nervous laughter he looked like he wanted to cry.

Yeah I noticed that... oh man... that's rough when Occulus guys are saying that behind closed doors.

This was what they said in public:

http://www.tomshardware.com/news/oculus-oculus-connect-vr-amd-nvidia,27729.html

Nvidia was making a big push for its new Maxwell class GTX980 GPUs. David Coombes is Nvidia's Developer Programs Manager, and he highlighted its special measures to dramatically reduce latency in VR applications with HMDs like the Oculus Rift.

According to Nvidia, its VR-optimized Maxwell GPUs can reduce latency by as much as 24ms compared to earlier GPU generations. This is a major deal when you figure that good VR experiences are aiming for 20ms or less of latency, and current Nvidia GPUs are part of a pipeline that is 57ms total.

57 - 24 = 33ms.

According to Tom Forsyth, Oculus VR's Software Architect, this is a new enhancement for Nvidia GPUs that is already available in AMD graphics cards. In fact, some of these optimizations can be found in current Xbox One and Sony PlayStation 4 platforms as part of their respective SDK. Why? Because they are based on AMD GPUs.

 

[monstercameron]

Techreport is obviously ANTI-AMD, as has been stated and confirmed by the many AMD hardcore fans on the forums here, so for them to be reporting Nvidia-anything being "potentially catastrophic" actually means Nvidia is going out of business tomorrow. Good game Nvidia, nice knowing ya.

i dont know about all that...but he was surprisingly terse on the subject.

 

[3DVagabond]

Techreport is obviously ANTI-AMD, as has been stated and confirmed by the many AMD hardcore fans on the forums here, so for them to be reporting Nvidia-anything being "potentially catastrophic" actually means Nvidia is going out of business tomorrow. Good game Nvidia, nice knowing ya.

LOL!

 

[Silverforce11]

Basically what Kanter is saying is what NV have been saying to VR developers:

https://developer.nvidia.com/sites/...works/vr/GameWorks_VR_2015_Final_handouts.pdf (p31)

All our GPUs for the last several years do context switches at draw call boundaries. So when the GPU wants to switch contexts, it has to wait for the current draw call to finish first.

So, even with timewarp being on a high-priority context, it’s possible for it to get stuck behind a longrunning draw call on a normal context. For instance, if your game submits a single draw call that happens to take 5 ms, then async timewarp might get stuck behind it, potentially causing it to miss vsync and cause a visible hitch.

Their hardware just cannot do graphics + compute together in parallel. It's not like GCN's hardware ACEs which can operate out-of-order, bypassing graphics traffic.

The question is why did they claim they could, saying they support it? Support doesn't infer all these software work-arounds that are gimmick to a point where Occulus guys are calling it a catastrophe.

 

[VulgarDisplay]

Basically what Kanter is saying is what NV have been saying to VR developers:

https://developer.nvidia.com/sites/...works/vr/GameWorks_VR_2015_Final_handouts.pdf (p31)



Their hardware just cannot do graphics + compute together in parallel. It's not like GCN's hardware ACEs which can operate out-of-order, bypassing graphics traffic.

The question is why did they claim they could, saying they support it? Support doesn't infer all these software work-arounds that are gimmick to a point where Occulus guys are calling it a catastrophe.

To sell gpu's to people stuck in the 28nm drought expecting high end GPU's released now to work well with DX12.

 

[VR Enthusiast]

The question is why did they claim they could, saying they support it? Support doesn't infer all these software work-arounds that are gimmick to a point where Occulus guys are calling it a catastrophe.

It's for the same reason why they said they have DX 12_1 and the same reason why they sold the GTX 970 with 0.5GB of fake VRAM. The press gives them an easy ride and they seem to have very forgiving customers.

Look at that toms article you linked from last year. It's really all about Nvidia then a little bit about AMD at the end, worded like AMD wasn't totally out of the VR fight. There won't be an article pointing out the opposite though.

VR matters to me. What I care about is VR done properly. Oculus should be ashamed of themselves - they are risking their profits and possibly the whole VR industry because they won't come out and just say that Nvidia can't do VR properly. Why are they telling experts like Kanter one thing yet still recommending the GTX 970 for VR? I hope Nvidia is paying them really well for their silence.

 

[monstercameron]

It's for the same reason why they said they have DX 12_1 and the same reason why they sold the GTX 970 with 0.5GB of fake VRAM. The press gives them an easy ride and they seem to have very forgiving customers.

Look at that toms article you linked from last year. It's really all about Nvidia then a little bit about AMD at the end, worded like AMD wasn't totally out of the VR fight. There won't be an article pointing out the opposite though.

VR matters to me. What I care about is VR done properly. Oculus should be ashamed of themselves - they are risking their profits and possibly the whole VR industry because they won't come out and just say that Nvidia can't do VR properly. Why are they telling experts like Kanter one thing yet still recommending the GTX 970 for VR? I hope Nvidia is paying them really well for their silence.

vr is subjective, especially the effects of higher latency.

 

[Vesku]

vr is subjective, especially the effects of higher latency.

Sure but Carmack says latency is a very big deal since it doesn't take much to hurt the experience for most users, which is why Oculus have an aggressive latency target.