Oxide: Nvidia GPU's do not support DX12 Asynchronous Compute/Shaders.
- Thread starter Arkanius
- Start date
SneakyStephan
Banned
I know the water under the level bit was debunked, but how about the concrete barriers that had 76778767777343973 polygons to render a flat surface?
Dictator93
Member
Though to be honest, if anyone is trying to offload a 980 Ti on the cheap (Europe) tell me your price
I want to see more Beyond3D benchmarks, those guys are really trying to test both architectures and trying to get some results out of it.
Can you link me the "bench"? I could test and post as well if you would like (Titan X).
Phong tessellation costs very little in comparison to other types (which is why it has that pattern stretch over it) and those images were taken in debug mode which ALSO disables Tesselation distance scaling and LODs. Those blocks only tessellate heavily (using the cheapest tessellation mode) when you are right right next to them. It is a non-issue. The best way to test how little performance it costs is by loading up the game on an older AMD card and toggling tesselation cvars in the console. It is easy to see how little it costs.
Sadly people confused the game running poorly in DX11 as lying with tesselation. Rather than looking at the super expensive post processing and pixel shaders.
Compute shaders, yes. You have to look at them talking about it in context of dx11 hardware. The code dealing with it could still be structured for serial execution.
Foutrologue
Banned
Compute shaders ≠ Async Compute/sharers
Compute sharers have been used for a while now. And has the blog post is saying; they are aiming for DX11 level hardware.
frontieruk
Member
Well we'll see a victory for NV under dx12 then
and the drama is on the other foot until the next DX12 game which will probably replicate that result, I'm not sure there is a situation outside this case that will see AMD ahead, I do think that AMD owners will just be grateful to get away from DX11 overheadsDictator93
Member
I was under the impression the game is DX12 only currently? And wouldnt AMD cards be better at a default compute (non-async)/serial one anyway?`
Foutrologue
Banned
Exactly. Water under level has always been part of the CryEngine since day one and didn't impact performance in-game (AFAIK is can now be disabled in the latest version). All the other crap (concrete slabs, pavements, brick walls..) was intentional to crush performance on Radeons (as seen by Crysis 3 where everything is toned down to acceptable levels with still good quality)
Dictator93
Member
No it wasnt. Water under the level never rendered in view. That is not how occlusion culling works in the engine. And you can go in the game yourself with some 5870 or whatever and test how little tesselation costs. It is seriously cheap.
Do you have an AMD GPU now? You can totally try it out with a CVAR to see how little it effects overall framerate.
Sure, but the game has been in development for a while now. You have to wonder how much of it is dx11 code transitioned to dx12.
Async compute would have to give AMD a 30-40% advantage in most DX12 games for me to care...
Typical 980 Ti OC is around 20% faster than OC'd Fury at 1440p, so, it may only really put them on par for me. I still play a lot of DX11 games and so performance in those is still important. I'll trade 10% less performance in DX12 if that's what it ends up being for that.
Typical 980 Ti OC is around 20% faster than OC'd Fury at 1440p, so, it may only really put them on par for me. I still play a lot of DX11 games and so performance in those is still important. I'll trade 10% less performance in DX12 if that's what it ends up being for that.
Foutrologue
Banned
I've used the CryEngine SDK on several occasions and know how its rendering works. Water under the level has never impacted performance. As a matter of fact the first thing you are presented with when you launch a new project is endless water on which you build your level on top (yeah it sounds ridiculous).
And yes ironically I still have a 5870 on which I ran C2 back then. Simply looking at a tessellated brick wall or concrete slab made the FPS tank into single digits.
frontieruk
Member
Its Windows 10 only, but it doesn't stop them from targeting DX11 hardware being the bulk of the market of machines upgrading, we kind of forget here not everyone splurges on a new card every 6mnths or buys a console day one.
serversurfer
Member
That's closer to my understanding of how this stuff works. Basically, you've got turnstile-style access to a fixed pool of resources; the various math units on the GPU, each with its own specialty. So think of it like loading a roller coaster. Every cycle, the system hangs the next rendering job on the GPU, occupying some or all of those specialized units. These jobs are the people who paid for VIP passes. Then the system looks at the math units that haven't been assigned jobs, compares that to the 64 jobs waiting at their respective turnstile all managed by eight line attendants and lets in whatever punters best fill the remaining seats before dispatching the train.
It sounds like NV do something similar, but instead of filling empty seats every cycle with jobs from the 31 compute queues, they actually alternate job types, pulling a job from the render queue on even cycles and a job (more?) from the compute queues on the odd cycles. Then they're saying, "Well, at the end of the day, everybody gets to ride." While it's true they're seamlessly pulling jobs from both queue types, because they can't pull from multiple queue types simultaneously, they're not actually doing much to increase utilization. Any math unit not used in a given render operation remains idle; it just gets used on the following cycle. I'm assuming they'd at least be able to pull from all 31 queues on the compute cycle to attempt to fully saturate the math units, but they'd still have a lot of idle units on the render cycle.
Is this a trick question because adding async to the mix "just" increases your peak utilization? It will have empty spaces in its rendering pipeline that need filling, just like any other GPU, if that's what you're asking.
icecold1983
Member
not according to what we know and rumors
Actually Unreal Engine 4 supports async compute on XB1, and I think it was lionhead who implemented it in unreal engine. There was also a gameplay video where they talked about it, I believe it was in the context of the pc version.
Yup, here you go.
https://docs.unrealengine.com/lates...ing/ShaderDevelopment/AsyncCompute/index.html
Thanks. However, they only mention Xbox One, and the documentation talks about it being disabled for PC in dx11.1 (possibly means that the documentation hasn't been updated).
Yeah, it says it's not supported in d3d11.1, because it literally isn't part of d3d 11.
Of course. There is no indication that it will be enabled for PC on dx12. Given that this documentation is tagged for UE 4.9 which was just released probably means it isn't available for PC. That is, if we're going by documentation, which we all know isn't always the most reliable thing.
serversurfer
Member
Why would it be available on Bone but not PS4?
Because the code was contributed by Lionhead.
Probably because lionhead probably did all the legwork here while coming up with a solution tailored for the xbox one for their game. It also says that epic intergrated it and will work on making it multiplatform if im interpreting that last paragraph correctly.
serversurfer
Member
Ah, thanks. <3
BennyBlanco
aka IMurRIVAL69
This could be your best work yet.
frontieruk
Member
You'd expect that of they've added the feature set for X1 they'd implement on PC though, wasn't that the point of windows 10 code once and deploy to many devices.
Should be 100% chance. It is part of DX12 after all.
The chance of seeing history repeat itself, Maxwell getting Kepler-ed in favor of Pascal, should be fun.
I wouldn't be so sure of it.
Pascal was supposed to be just Maxwell with HBM. Don't forget that for Nvidia Maxwell is full DX12 compliant
What they say and what actually happens can be totally different things.
Also guys correct me if I'm wrong. But aren't those homebrew tests proving that Maxwell does have Async Compusing but limited to 31 queues ? The test goes by submitting 128 threads. GCN takes them all with the same latency but Maxwell can only take 31 (It has 31+1) queues before processing the next load.
The only thing it shows is that it has a lower latency for low queues processing and a equivalent latency compared to AMD.
They do less but more quickly. GCN does more but slower
The only thing it shows is that it has a lower latency for low queues processing and a equivalent latency compared to AMD.
They do less but more quickly. GCN does more but slower
icecold1983
Member
seem unlikely because i think at this point its pretty cut and dry that nvidias hardware is poorly suited for async compute, but itd be interesting to hear an official response from nvidia.
To me it looks like the tests aren't showing that at all. Nvidia cards are finishing the graphics + compute test in the same time as both the graphics test and compute test combined, while amd cards are finishing the graphics + compute test in the same time as the compute test alone.
Yep. Sebs said he expects some games to hit furmark levels of load. Maybe only on the xb1 though, I've heard the ps4 already gets pretty loud.
serversurfer
Member
That would seem to confirm the idea NV are simply alternating between the two job types rather than blending them.
It reminds me of intel ht working better on i7 better than amd fx cmt because of overall IPC superiority from intel. And as long as Nvidia can tweak the implementation of async. to be decent, around 70% of same in class, it doesn't matter how they get there unless you want more transparency from their marketing guys.
Hopefully everyone gets a win. If things as is stand then it's pretty loose with dx12 compliance.
The performance of a video game is defined by more than just frames per second or frametimes. Asynchronous compute allows for higher throughput at lower latencies which easily makes it one of the most important features for VR gaming. Remember the beginning of this gen when Mark Cerny explained again and again the importance of async compute for the future of video games? That was before Morpheus was announced. Two years later it all makes sense.
Coulomb_Barrier
Member
Glad I didn't fork out a ludicrous £550 for a 980 Ti that will struggle with demanding DX12 games in the next couple years. As if I spent that kind of money, I'd expect to not have to upgrade for 2+ years. But then again, my buying habits are different from most who like to upgrade yearly.
What I've done instead is fork out a large sum for an EVGA 980 that will probably last less time
Anyway, these are all good cards for a while yet at least.
What I've done instead is fork out a large sum for an EVGA 980 that will probably last less time
Anyway, these are all good cards for a while yet at least.
MutonCommander
Banned
What about normal games? Should we expect better performance in them? Will the PS4 perform as well as a higher end PC graphics card?
frontieruk
Member
Glad I didn't fork out a ludicrous £550 for a 980 Ti that will struggle with demanding DX12 games in the next couple years. As if I spent that kind of money, I'd expect to not have to upgrade for 2+ years. But then again, my buying habits are different from most who like to upgrade yearly.
What I've done instead is fork out a large sum for an EVGA 980 that will probably last less time
Anyway, these are all good cards for a while yet at least.
It depends on the volume of async calls being used, as said earlier in the thread devs aren't going to gimp their games on the larger 80% of hardware, though even if they did you'd probably see comparable performance under dx11 without the effects async would be used for anyway, this is getting silly, its being pushed hard on consoles because the CPUs are weak, we'll see devs like Dice play with async as they are already leveraging it but to the levels that this fuss is over is really doubtful.
Normal games? All games are normal unless you're referring to dudebro AAA.
No, they'll leverage for more pretty than performance.
No.
Dictator93
Member
Turin Turambar
Member
It's also more efficient that way, otherwise you have "wasted silicon", with idle times here and there.
Nvidia itself specifically told the developer to turn off the feature for their card. If it were just a bug they could like fix it instead.
thelastword
Banned
Good post, this is pretty much what it is, and the benefits derived from the PS4 architecture. People must know that async compute is being used on PS4, yes, but still marginally. That sort of stuff takes time.
Although ISS used it, I'm going to say that Uncharted 4 will be the first game to use the feature (at least on the level that Cerny alluded to) and even moreso by first party Sony games going forward, from March 2016 when that game launches and beyond. I believe that's when we will witness some consistent results in games relative to GPGPU computing.
Also, good contributions by ServerSurfer and Arkanius..........
So is this false advertising then?: -
https://forums.geforce.com/default/...s/about-nvidia-and-dx12/post/4656838/#4656838
https://forums.geforce.com/default/...s/about-nvidia-and-dx12/post/4656838/#4656838
So is this false advertising then?: -
https://forums.geforce.com/default/...s/about-nvidia-and-dx12/post/4656838/#4656838
No, because Nvidia supports it.
Badly, but it is there.
Dictator93
Member
We, IMO, still really have a hard idea understanding what is happening. Has that benchmark gotten better yet?