D. Kanter On AS - 'Oculus: Preemption For Context Switches Best On AMD, Good On Intel, Possibly Catastrophic For NVIDIA'

[HKZeroFive]

The video in question - Please fast forward to 1:21:35

 

 

David Kanter (Microprocessor Analyst): I've been told by folks at Oculus, the best preemption support for context switches was AMD by far. Intel was pretty good and NVIDIA was possibly catastrophic.

 

The real issue is, if you have a graphics shader running, you need to let it finish. It could take you a long time, it could take you over 16 milliseconds to finish.

 

I've heard from engineers...graphics shaders that took a couple seconds to run. Obviously something was totally busted. This is really great for your PowerPoint slideshow.

 

NVIDIA to their credit, has been very open and honest about this on how to tune for Oculus Rift, is that you have to be super careful cause you can miss a frame boundary because the preemption is not particularly low latency.

 

This is a pretty significant problem for VR in particular, and I know this was obvious (about AMD having an advantage over NVIDIA because of their async compute capibilities), but AMD's is TrueAudio seeming to be the better way to go for VR.

 

I don't like the political stance Oculus has on the issue, I feel like they should be more open on 'what is best, what is not good' and so forth.

 

This is probably going to stir up a mild 'crap hitting the fan' storm but this is just another example of why NVIDIA needs to focus on implementing a better async compute in the future. The tide is indeed seemingly to be turning in AMD's favour, and NVIDIA, even with their 80% market share, is not immune to the market's flexibility and adaptive nature.

 

But man, you've got to be doing something wrong if Intel has better graphics support than you...

 

TL;DR

NVIDIA is getting bit in the bum because they didn't predict the importance of AS when making Maxwell. AMD played the future game and it looks like it's pulling off at the moment. According to Oculus, Intel > NVIDIA.

[Tim Drake]

Tl;DR?

[Me1z]

Tl;DR?

 

Nvidia has been focusing so much on efficiency and DX11 performance that they dropped the ball when it comes to predicting the adoption of new technologies. There's no good reason for it besides it being a blunder on Nvidia's part, but AMD has significant advantages when it comes to VR, DX12, and the direction the gaming industry is moving as a whole. Pascal will likely address these issues though.

[Tim Drake]

Nvidia has been focusing so much on efficiency and DX11 performance that they dropped the ball when it comes to predicting the adoption of new technologies. There's no good reason for it besides it being a blunder on Nvidia's part, but AMD has significant advantages when it comes to VR, DX12, and the direction the gaming industry is moving as a whole. Pascal will likely address these issues though.

Wooo, AMD do good!

[Zah]

Wooo, AMD do good!

Now the hard part is to keep that advantage.

[bcredeur97]

nvidia is usually faster at mobilizing though... they will likely overrun AMD in the near future.

[Clyne]

Nvidia has been focusing so much on efficiency and DX11 performance that they dropped the ball when it comes to predicting the adoption of new technologies. There's no good reason for it besides it being a blunder on Nvidia's part, but AMD has significant advantages when it comes to VR, DX12, and the direction the gaming industry is moving as a whole. Pascal will likely address these issues though.

It's like 3dfx all over again. They focused on Banshee and Avenger for 3+ years on glide/direct3d, once they were released, it was obsolete as technology was advancing from Glide/direct3d to DirectX. 

But i doubt Nvidia will end up in a situation like 3dfx......

[Paranoid Kami]

I'll probably be selling my 970 and buying an AMD card when the next generation hits. I didn't mind the gimped Vram at 3.5, since I never used that much and the Batman fiasco wasn't their fault, but I bought a new card specifically for DX12 and VR and now I'm hearing that it's completely broken in both of them. It's going to take quite a few years for me to ever trust Nvidia again after this.

[Notional]

It's like 3dfx all over again. They focused on Banshee and Avenger for 3+ years on glide/direct3d, once they were released, it was obsolete as technology was advancing from Glide/direct3d to DirectX. 

But i doubt Nvidia will end up in a situation like 3dfx......

 

True. The main question is though: Pascal has already been taped out and is in (pre)production, so has it already solved the issues of async compute? If not then Pascal is screwed too. Either way everyone with a Kepler (700 series) or a Maxwell (900 series) might be pretty screwed for VR and DX12 games. I guess the only saviour would be proprietary GameWorks titles that has been forced to not use async compute.

[Clyne]

True. The main question is though: Pascal has already been taped out and is in (pre)production, so has it already solved the issues of async compute? If not then Pascal is screwed too. Either way everyone with a Kepler (700 series) or a Maxwell (900 series) might be pretty screwed for VR and DX12 games. I guess the only saviour would be proprietary GameWorks titles that has been forced to not use async compute.

It's Nvidia, pretty sure they can make a solution like always.

With that 970 VRAM fiasco, nobody really cared for it and it still performed fine. 

[BuckGup]

Yeah AMD is ahead right now. Come 2016 and if ZEN isn't a CPU from the gods AMD is getting in some real trouble 

[Notional]

Yeah AMD is ahead right now. Come 2016 and if ZEN isn't a CPU from the gods AMD is getting in some real trouble 

 

In what way? Pascal has no pros compared to Greenland. On the contrary, is Pascal is as useless at async compute as Kepler/Maxwell, I think NVidia will have a very big problem next year.

[qwertywarrior]

maxwell's latency is 33.33ms

amd is i believe <20

[Derangel]

The whole async thing is interesting but I wonder where it will end up. We may see AMD get an advantage on VR for now but I wonder about DX12. People are making a lot of noise about Ashes (a game no one seemed to give two shits about before the benchmarks, but whatever) and it's engine. It's nice to see some tech that AMD really excels at, but I have a wrinkle to throw into that hole debate: Unreal Engine 4. Nvidia worked with Epic on development of UE4 to the point where the default physics engine is the CPU based version of PhysX. If history repeats itself UE4 will by far be the most dominate 3rd party engine for DX12 games. Games developed on UE4 may not use async as heavily. Consoles tend to throw some wrinkles into the mix as well since developers can take advantage of async when making console games. Either way I think people need to cool their jets a little bit and wait to see what happens when graphically intensive, designed for the API, DX12 games hit sometime next year. If AMD managed to plan that well and their GCN cards really do end up being amazing at DX12 that will be great and would definitely push the value of their current line-up up a few notches. At this point all of this stuff is an interesting discussion topic but not really big enough to say "Nvidia is doomed" or "AMD will win" or whatever else.

 

On the subject of VR itself: As much as I'm really looking forward to getting a VR headset it'll be years before most enthusiasts are worried about which brand is better for them, much less the mass market. By that time all the dust will have settled and both companies will have cards out designed to handle VR.

[BuckGup]

In what way? Pascal has no pros compared to Greenland. On the contrary, is Pascal is as useless at async compute as Kepler/Maxwell, I think NVidia will have a very big problem next year.

Pascal will have incredible AI support for deep learning. Both will have extreme memory bandwith but both focus on a different piece they think will help them. It is true Nvidia has been just sitting there as they have had no competition as AMD is lacking in raw performance but has great features like async compute. We will just have to see if they can get their CPUs down.

[Zah]

The whole async thing is interesting but I wonder where it will end up. We may see AMD get an advantage on VR for now but I wonder about DX12. People are making a lot of noise about Ashes (a game no one seemed to give two shits about before the benchmarks, but whatever) and it's engine. It's nice to see some tech that AMD really excels at, but I have a wrinkle to throw into that hole debate: Unreal Engine 4. Nvidia worked with Epic on development of UE4 to the point where the default physics engine is the CPU based version of PhysX. If history repeats itself UE4 will by far be the most dominate 3rd party engine for DX12 games. Games developed on UE4 may not use async as heavily. Consoles tend to throw some wrinkles into the mix as well since developers can take advantage of async when making console games. Either way I think people need to cool their jets a little bit and wait to see what happens when graphically intensive, designed for the API, DX12 games hit sometime next year. If AMD managed to plan that well and their GCN cards really do end up being amazing at DX12 that will be great and would definitely push the value of their current line-up up a few notches. At this point all of this stuff is an interesting discussion topic but not really big enough to say "Nvidia is doomed" or "AMD will win" or whatever else.

 

On the subject of VR itself: As much as I'm really looking forward to getting a VR headset it'll be years before most enthusiasts are worried about which brand is better for them, much less the mass market. By that time all the dust will have settled and both companies will have cards out designed to handle VR.

This is quite true.  Async computer has to specifically be programmed for, and although Unreal 4 can do it, it's not recommended for console games, don't know about for desktop applications in this sense.

 

VR, AMD will have the upper hand, though, if nvidia's Async warp tech doesn't match AMD's shaders.  However, consumer VR isn't out yet, so laypeople do not have access, and we can't always trust devs and journalists.

[spartaman64]

Pascal will have incredible AI support for deep learning. Both will have extreme memory bandwith but both focus on a different piece they think will help them. It is true Nvidia has been just sitting there as they have had no competition as AMD is lacking in raw performance but has great features like async compute. We will just have to see if they can get their CPUs down.

no competition? amd's line up beats nvidia's line up at almost every price range except for the 980 ti and fury x. the r9 390 is faster than the 970, the r9 380 beats the 960

[That Norwegian Guy]

Pascal will likely address these issues though.

 

Doubtfully, Pascal was on Nv's roadmap for 2nd quarter 2016, that means the actual design is well beyond finished. Either they scrap Pascal and rush to Volta or release Pascal with the same old issues. I think Nvidia will somewhat arrogantly push Pascal anyway, expecting their existing market position to be enough to sell well.

[ACatWithThumbs]

 

TL;DR

NVIDIA is getting bit in the bum because they didn't predict the importance of AS when making Maxwell. AMD played the future game and it looks like it's pulling off at the moment. According to Oculus, Intel > NVIDIA.

Ohh they knew it, Maxwell was just not built as it was supposed to due to 20nm not being ready at the time and them needing to cut features to put the bigger SMX on 28nm.

Maxwell was supposed to have Double Precision and Virtual Unified Memory with probably around 5000 Cuda Cores and I bet A-Sync was also planned.

That was the reason I didn't buy Maxwell in the first place because it was nothing like on their roadmap and extremely slow compared to what it was supposed to be.

 

Doubtfully, Pascal was on Nv's roadmap for 2nd quarter 2016, that means the actual design is well beyond finished. Either they scrap Pascal and rush to Volta or release Pascal with the same old issues. I think Nvidia will somewhat arrogantly push Pascal anyway, expecting their existing market position to be enough to sell well.

 

Pascal won't have any of those issues as Maxwell was a mess due to 28nm and I'm 100% sure A-Sync Compute was supposed to be in Maxwell.

 

[BuckGup]

no competition? amd's line up beats nvidia's line up at almost every price range except for the 980 ti and fury x. the r9 390 is faster than the 970, the r9 380 beats the 960

What about the GTX980? Or GTX950? Or Titan X? Enthusiasts want the best of the best they can get at the moment and that is the GTX980ti so they turn to it rather than AMD.

[Notional]

Nvidia why you so scummy?

 

Because people buy their overpriced stuff anyways. The 970 is the most popular card of the high end gaming world. The one card, that ripped off people with 3.5GB of vram. So the answer is because we let them.

 

Pascal will have incredible AI support for deep learning. Both will have extreme memory bandwith but both focus on a different piece they think will help them. It is true Nvidia has been just sitting there as they have had no competition as AMD is lacking in raw performance but has great features like async compute. We will just have to see if they can get their CPUs down.

 

In FLOPS AMD is actually ahead of NVidia. In gaming, AMD beats NVidia in every price point except the extreme high end. That will change with DX12 and possibly Vulkan. NVidia's general performance gain has been from multithreaded low overhead drivers. Those are rendered irrelevant in DX12.

 

This is quite true.  Async computer has to specifically be programmed for, and although Unreal 4 can do it, it's not recommended for console games, don't know about for desktop applications in this sense.

 

VR, AMD will have the upper hand, though, if nvidia's Async warp tech doesn't match AMD's shaders.  However, consumer VR isn't out yet, so laypeople do not have access, and we can't always trust devs and journalists.

 

Actually it is recommended for consoles as they are GCN based. The guy from Oxide games even said, that some devs are reaching plus 30% performance with async compute on the consoles. So we might see a lot of games on PC reusing the async compute code that the console versions are using.

 

Pascal won't have any of those issues as Maxwell was a mess due to 28nm and I'm 100% sure A-Sync Compute was supposed to be in Maxwell.

 

We don't know anything about Pascal tbh. Pascal was taped out before the Ashes of the Singularity benches hit the web, so either NVidia already fixed it before, or Pascal will suffer too. We just don't know. Personally I don't think NVidia would be dumb enough not to support async compute in Pascal, but with all the BS NVidia is pulling lately, I really don't know.

[spartaman64]

What about the GTX980? Or GTX950? Or Titan X? Enthusiasts want the best of the best they can get at the moment and that is the GTX980ti so they turn to it rather than AMD.

but most people will buy midrange cards so nvidia's massive market share is more likely because of fanboys and people who dont know a lot about computers. and the r9 fury beats the 980 albeit at a slightly higher price of about 50 dollars and the r9 390x is only slightly slower at 100 dollars less. and we will see when the gtx 950 launches amd will probably cut the price of the r7 370

[HKZeroFive]

What about the GTX980? Or GTX950? Or Titan X? Enthusiasts want the best of the best they can get at the moment and that is the GTX980ti so they turn to it rather than AMD.

The Titan X is irrelevant in gaming now, since NVIDIA basically killed it off by introducing the 980Ti as a pre-emptive measure to counter AMD's release of the Fury X.

As for the 980, I was kinda hoping the Nano would handle that but $650 is way too much to ask, even if they're aiming for the premium mini-ITX market.

I found this small comic that kind of sums up NVIDIA's situation:

Classic NVIDIA

[Derangel]

Because people buy their overpriced stuff anyways. The 970 is the most popular card of the high end gaming world. The one card, that ripped off people with 3.5GB of vram. So the answer is because we let them.

 

No, it's the most popular because it's the best price-performance card on the market (Though, good arguments can be made for the 390 these days). Nvidia's lies and PR bullshit don't change that. The 970 didn't magically perform worse after the VRAM stuff was found out. No one was ripped off because the performance of the card was widely reported at launch. The 970 is what it is regardless of the shit Nvidia pulled. Neither AMD nor Nvidia are "good" companies.

[Notional]

No, it's the most popular because it's the best price-performance card on the market (Though, good arguments can be made for the 390 these days). Nvidia's lies and PR bullshit don't change that. The 970 didn't magically perform worse after the VRAM stuff was found out. No one was ripped off because the performance of the card was widely reported at launch. The 970 is what it is regardless of the shit Nvidia pulled. Neither AMD nor Nvidia are "good" companies.

 

The problem is that people generally don't buy a card for one month. They buy them to last for years, which is why future proof is not only a valid point but a necessary one at that. When the 970 was reviewed, it did not benchmark any vram heavy games. In fact most games at the point was not vram intensive, as most games were still held back by the old generation of consoles. The picture got a little different when Dying Light was released.

Both AMD and NVidia are going for more than 4GB on their Pascal/Greenland cards for a reason: It will be used in upcoming games.

 

Granted both are companies and do what they need to sell stuff. But NVidia is by far the most anti consumer company of the two, with planned obsolescence, subpar hardware and proprietary everything resulting in price premiums and vendor lock in.