[Updated] Oxide responds to AotS Conspiracies, Maxwell Has No Native Support For DX12 Asynchronous Compute

[Briggsy]

might also want to look into this: https://www.reddit.com/r/nvidia/comments/3j5e9b/analysis_async_compute_is_it_true_nvidia_cant_do/

source: https://forum.beyond3d.com/posts/1869058/

 

The first test reveals just how parallel AMD deals with compute, vs Nvidia's serial approach. Not surprising.

 

The second test should be taken with a large grain of salt, as it is a controlled and predictable async compute test, which Nvidia's scheduler can handle easily. Also, not surprising.

 

The real test is when used in real time, with a game engine that is throwing unpredictable and random compute and graphic tasks at the graphic card, and the resulting latency.

[zMeul]

snip

no, the issue at hand was if Maxwell can or can't do async compute - and that's it

how it's handled, that's another story and a different topic

to read from Unreal Engine 4.9 documentation: https://docs.unrealengine.com/latest/INT/Programming/Rendering/ShaderDevelopment/AsyncCompute/index.html

AsyncCompute should be used with caution as it can cause more unpredicatble performance and requires more coding effort for synchromization.

AMD's PR machine is in full swing and it's gaining traction because they know nVidia doesn't respond to situations like this, unless it gets very dramatic

a lot of idiots already started returning their GTX video cards because of this

[Briggsy]

no, the issue at hand was if Maxwell can or can't do async compute - and that's it

how it's handled, that's another story and a different topic

to read from Unreal Engine 4.9 documentation: https://docs.unrealengine.com/latest/INT/Programming/Rendering/ShaderDevelopment/AsyncCompute/index.html

 

 

I don't think (at least I'm not arguing) that Nvidia can't do async compute. the problem-as you say-is how its handled with Maxwell. Maxwell seudo-supports async compute, but suffers latency penalties when used in gaming. The discussion was always about how well Maxwell could handle async compute in a real time gaming environment. 

[zMeul]

Maxwell seudo-supports async compute

no, that's the wrong way to think - does Maxwell support async compute? yes! not pseudo nothing

Oxide's benchmark was clearly programmed with async path to be benefic towards GCN and not towards Maxwell - it is a AMD sponsored developer after all; if nVidia would be the sponsor, then the table would be reversed .. no question

[VagabondWraith]

............

 

AMD's PR machine is in full swing and it's gaining traction because they know nVidia doesn't respond to situations like this, unless it gets very dramatic

a lot of idiots already started returning their GTX video cards because of thisI'll

I'll laugh my ass off if these "benchmarks" prove to be nothing more than AMD propaganda. In return making people returning their cards even more of a dumbass than they already were.

[MageTank]

I'll laugh my ass off if these "benchmarks" prove to be nothing more than AMD propaganda. In return making people returning their cards even more of a dumbass than they already were.

Hey, let people return their GPU's. I'll snag a used GTX 980 Ti for cheap.

[zMeul]

my guess is Oxyde coded the async shaders with OpenCL, and no surprise .. nVidia doesn't do well with OpenCL, but it does very well with CUDA

ultimately their decision to slack on OpenCL development will bite them in the ass

[HalGameGuru]

my guess is Oxyde coded the async shaders with OpenCL, and no surprise .. nVidia doesn't do well with OpenCL, but it does very well with CUDA

ultimately their decision to slack on OpenCL development will bite them in the ass

Don't quote ME on this but I could have sworn someone in the OC.net thread was talking about DirectCompute in DX12. But, I'd have to reread thru 150 pages to find it and if it was making mention of this specific instance.

[Zah]

my guess is Oxyde coded the async shaders with OpenCL, and no surprise .. nVidia doesn't do well with OpenCL, but it does very well with CUDA

ultimately their decision to slack on OpenCL development will bite them in the ass

I'm not a dev, but is it possible to make a path to use OpenCL for AMD drivers, CUDA for NVidia?

[zMeul]

I'm not a dev, but is it possible to make a path to use OpenCL for AMD drivers, CUDA for NVidia?

yes it is, but will AMD be OK with it?

the real problem - it's a whole lot of work to be done writing two different paths because ultimately it can end in creating two game executables, one for nVidia other for AMD

[zMeul]

Don't quote ME on this but I could have sworn someone in the OC.net thread was talking about DirectCompute in DX12. But, I'd have to reread thru 150 pages to find it and if it was making mention of this specific instance.

direct compute exists in DX since DX10, that's not the issue

my question: is Async Shaders a core component of DX12 (meaning mandatory) or is one of the tiered optional ones - I tried to google it, but because of this "drama" there's a lot of noise

[xAcid9]

how it's handled, that's another story and a different topic

to read from Unreal Engine 4.9 documentation: https://docs.unrealengine.com/latest/INT/Programming/Rendering/ShaderDevelopment/AsyncCompute/index.html

Only for Nvidia gpu it seem, Kollock even mentioned that hell break loose when they try to enable async compute for Maxwell.

[HalGameGuru]

direct compute exists in DX since DX10, that's not the issue

my question: is Async Shaders a core component of DX12 (meaning mandatory) or is one of the tiered optional ones - I tried to google it, but because of this "drama" there's a lot of noise

It would appear it is one of the mandatory functions of Tier 2 or better DX12 support. Intel's GPU's at the time of the render did not support it at all, and were Tier 1, and the nVidia GPU's that did not support it at all were Tier 1, but the Tier 2's and 3's supported it.

 

But, then there are conflicting reports, some say certain features ARE supported or are NOT supported by the same GPU's. Unless and until Microsoft validates for sure the individual GPU's we may not know exact tiers.

[PillowSmoke]

how it's handled, that's another story and a different topic

to read from Unreal Engine 4.9 documentation: https://docs.unrealengine.com/latest/INT/Programming/Rendering/ShaderDevelopment/AsyncCompute/index.html

Only for Nvidia gpu it seem, Kollock even mentioned that hell break loose when they try to enable async compute for Maxwell.

 

nice gif of taeyeon

[Briggsy]

direct compute exists in DX since DX10, that's not the issue

imagegifpng.jpg

my question: is Async Shaders a core component of DX12 (meaning mandatory) or is one of the tiered optional ones - I tried to google it, but because of this "drama" there's a lot of noise

 

it can be disabled, so its not a mandatory feature.

 

It a lengthy read but here's microsoft's official documentation on it: https://msdn.microsoft.com/en-us/library/windows/desktop/dn899217(v=vs.85).aspx

[HalGameGuru]

it can be disabled, so clearly its not a mandatory feature.

 

It a lengthy read but here's microsoft's official documentation on it: https://msdn.microsoft.com/en-us/library/windows/desktop/dn899217(v=vs.85).aspx

Not mandatory to make use of DX12, but an integral part OF DX12, and apparently required for a certain tier level of support for Feature Level 0 in DX12.

[Sammael]

no, that's the wrong way to think - does Maxwell support async compute? yes! not pseudo nothing

Oxide's benchmark was clearly programmed with async path to be benefic towards GCN and not towards Maxwell - it is a AMD sponsored developer after all; if nVidia would be the sponsor, then the table would be reversed .. no question

 

 

Maxwell supporting async compute, depending on how it works in practice, could be worse than saying the 970 has 4GB of video memory (the end - without any qualifications).

 

Why?  Because up to the 3.5 GB mark, the 970 performed as one might expect and people did not get hit hard for what was left unsaid about how the memory worked.   With Asynch compute, not being able to perform compute and general graphics tasks in parallel means that your cards don't use asynchronous compute in the ways that give the desired benefits.  That's a problem.  You can say that games that design their engines to utilize asynchronous  compute are being paid off by amd, but that's like saying any engine developer that actually makes use of a core dx12 feature is being paid off by a specific gpu company.  That is part of the god damn spec.  If nvidia cards can't deal with it as well, then they need to take their lumps just like amd did when dx11 cpu performance.

 

 

I've read that Unreal Engine 4 does not rely as heavily on asynch compute, in that case I very much do believe that part of the reason why is that they ARE sponsored up the ass by nvidia and gameworks tie ins, and if using the more modern style of rendering - asynch compute performs worse for nvidia, then they may as well not focus so heavily on that.  Stick to more serialized workloads, where their sugar daddy nvidia performs better.

[Briggsy]

Maxwell supporting async compute, depending on how it works in practice, could be worse than saying the 970 has 4GB of video memory (the end - without any qualifications).

 

Why?  Because up to the 3.5 GB mark, the 970 performed as one might expect and people did not get hit hard for what was left unsaid about how the memory worked.   With Asynch compute, not being able to perform compute and general graphics tasks in parallel means that your cards don't use asynchronous compute in the ways that give the desired benefits.  That's a problem.  You can say that games that design their engines to utilize asynchronous  compute are being paid off by amd, but that's like saying any engine developer that actually makes use of a core dx12 feature is being paid off by a specific gpu company.  That is part of the god damn spec.  If nvidia cards can't deal with it as well, then they need to take their lumps just like amd did when dx11 cpu performance.

 

 

I've read that Unreal Engine 4 does not rely as heavily on asynch compute, in that case I very much do believe that part of the reason why is that they ARE sponsored up the ass by nvidia and gameworks tie ins, and if using the more modern style of rendering - asynch compute performs worse for nvidia, then they may as well not focus so heavily on that.  Stick to more serialized workloads, where their sugar daddy nvidia performs better.

 

It will definitely be interesting with ARK dx12 patch (I own and play ARK, and there does appear to be a number of compute based effects in use). If they disable AC completely, it will only add confirmation that Nvidia can't properly execute graphic and compute tasks in parallel. If they enable it only for AMD, same thing. If they leave it enabled for Maxwell and there aren't any problems, its a win-win, and the industry can move forward. I don't see the last scenario playing out, but it would be nice. 

[Humbug]

It will definitely be interesting with ARK dx12 patch (I own and play ARK, and there does appear to be a number of compute based effects in use). If they disable AC completely, it will only add confirmation that Nvidia can't properly execute graphic and compute tasks in parallel. If they enable it only for AMD, same thing. If they leave it enabled for Maxwell and there aren't any problems, its a win-win, and the industry can move forward. I don't see the last scenario playing out, but it would be nice.

I'm guessing they will enable AC by default in DX12. Or they would be underutilizing not only AMD but upcoming Nvidia Pascal GPUs. They can specifically detect Maxwell / kepler GPUs and disable it if necessary.

[shymi]

no, the issue at hand was if Maxwell can or can't do async compute - and that's it

how it's handled, that's another story and a different topic

to read from Unreal Engine 4.9 documentation: https://docs.unrealengine.com/latest/INT/Programming/Rendering/ShaderDevelopment/AsyncCompute/index.html

AMD's PR machine is in full swing and it's gaining traction because they know nVidia doesn't respond to situations like this, unless it gets very dramatic

a lot of idiots already started returning their GTX video cards because of this

Just saw this in guru3d forums. Well, basically I'm pissed - "don't use parallel programming, because resource management is tougher". Can't add anything else.

 

About nVidia taking it like a man - yeah right. That's why they demanded async computing to be disabled by Oxide and telling developers through Unreal team not to use it, because it's harder to code. It's a big f**k up from the green team and they just keep adding oil to the fire with their PR stunts with developers. 

P.S. AMD also had it's share of s**t storm with rebranding, false claims for overclocking the Fury line and so on - so they are also not innocent...

[zMeul]

snip

it is not that simple at all

the main fault in all this is MS implementing different tiers of DX12 - and I suspect all 4 (MS, AMD, Intel and nVidia) of them had a hand in this

the problem is that none of the current hardware is supporting DX12 , fully; and this is biggest issue

if a developer creates a game using DX12 tier3, then the lower tiered GPUs will have to emulate the functions, or simply not work at all

Async Compute is there in nVidia's Maxwell, the TESLA accelerators use it currently

the big question is why Oxyde's benchmark can't utilize it - is it because Maxwell can't render async shaders and "normal" shaders in the same pipeline? or is it because Oxyde coded the game specifically for GCN

[Praesi]

So AMD equaled with NVidia in the first DX12 Title. Wow, NVidia you realy messed up not to be 20% ahead like all the Years bevore.

And since its far too soon to shitstorm this Topic, we all should go crazy about it.

[zMeul]

Not mandatory to make use of DX12, but an integral part OF DX12, and apparently required for a certain tier level of support for Feature Level 0 in DX12.

if it's "optional", then something smells fishy because Oxyde claimed nVidia asked them to disable it

but, if it's "optional" why Oxyde hasn't disabled it?!

[Ekpyrosis]

if it's "optional", then something smells fishy because Oxyde claimed nVidia asked them to disable it

but, if it's "optional" why Oxyde hasn't disabled it?!

Oxyde tried to use async compute at first, because the driver reports that the feature is available (otherwise they wouldn't have tried to use it in the first place), but perfomance was worse than without it, so they disabled it for Nvidia GPUs specifically to increase perfomance.

The thing that Nvidia asked them to disable wasn't async compute, they rather wanted them to change game settings like MSAA which would change the payload and effectively be a whole different benchmark, Oxyde refused to do that, it wasn't directly related to async compute.

[zMeul]

snip

I stand corrected, I remember it wrong