AMD Announces Exciting DirectX 12 Game Engine Developer Partnerships

[the54thvoid]

Yeah, about async compute ... it's super easy on GCN because it is perfectly happy to accept compute commands in the 3D queue. There is no penalty for mixing draw calls and compute commands in the 3D queue.
With Maxwell you have performance penalties from using compute commands concurrently with draw calls, so compute queues are mostly used to offload and execute compute commands in batch.
Essentially if you want to use async compute efficiently on nvidia, you gotta cleanly separate the render pipeline into batches and even consider including CUDA.dll to fully use high priority jobs and independent scheduling (with GK110 and later, CUDA bypasses the graphics command processor and is handled by a dedicated function unit in hardware which runs uncoupled from the regular compute or graphics engine. It even supports multiple asynchronous queues in hardware). It's a complete mess, and all detailed here: http://ext3h.makegames.de/DX12_Compute.html

Nice read. Sort of.

It's possible for a dev to work with CUDA to make async work then. That would require Nvidia to sponsor titles and help with coding for CUDA to prioritise the batches to suit the hardware. The article said that would mean worse case for AMD but good gains for Nvidia as the CUDA route allows the hardware to do async batches better. Vice versa is the hardware only solution as AMD has designed GCN for, which is worse case for Nvidia.

So, AMD can sponsor titles and Nvidia lose out or Nvidia can sponsor titles and AMD can lose out.

No change then!

 

[FordGT90Concept]

To use CUDA is to disable the 3D pipeline. NVIDIA cards require switching between the two. That doesn't go over well in games because the 3D pipeline is far more important.

GCN doesn't care if it is compute or 3D, it queues it into the same pipeline.

NVIDIA needs to fix it otherwise games just won't use it.

 

[Vayra86]

Getting worked up about perf differences with current gen cards in DX12 mode:

1. Pointless
2. Predictable
3. Rather naive

I have underlined this with the release of maxwell and the PR around current gen in terms of being dx12 ready, every single time: NONE of the current gen cards are really ready and buying into them and expecting the opposite is pretty short sighted. We all knew big changes were close and here they are. New arch/next gen gpu will be ready for it and current gen is already bottlenecked in many other ways, most notably CPU load and VRAM.

Stop worrying for nothing because it only underlines how uneducated your purchase has been. And it displays a lack of insight in the way the industry works.

Next gen selling points are going to be bigger leaps in DX12 performance; thats how they get us to buy new product.

 

[Xzibit]

Straight consoles ports will benefit AMD. Nvidia would have to sponsor to null (DX12 to DX11) and implement its beneficial code like the new Tomb Raider.

 

[the54thvoid]

Straight consoles ports will benefit AMD. Nvidia would have to sponsor to null (DX12 to DX11) and implement its beneficial code like the new Tomb Raider.

Straight console ports are abominations. Frame locked, key bindings screwed and PC graphics settings not there.

 

[TheHunter]

My system runs all my games at max possible settings at all times, it's this specific shit that always fucks up everything. And I'm not even blaming AMD here. They've done DX12 right, it's NVIDIA that was lazy. But I love Deus Ex franchise, that's why I'm worrying.

Then again, Deus Ex Human Revolution looked amazing so if I get that level of graphics I'm fine with it anyway. So yeah, chilling...

DeusEx human revolution was a AMD evolved game too and it ran perfectly fine on Nvidia.. actually it ran much better then amd later.
I have no doubt it will be the same now.

As for these games list.. nothing but yawn fest.. only this new DeusEx looks interesting, rest not worth mentioning.

 

[FYFI13]

Wake me up when devs switch to Vulkan and drop DX completely.

DirectX 12 performs better than Vulkan, according to first tests BUT... Vulkan works on most available platforms and only because of that DirectX should be ditched for good!

 

[BiggieShady]

To use CUDA is to disable the 3D pipeline. NVIDIA cards require switching between the two. That doesn't go over well in games because the 3D pipeline is far more important.

It is doable, Just Cause 2 had a water simulation done in pure CUDA.
Quoted from the article:

Ask your personal Nvidia engineer for how to share GPU side buffers between DX12 and CUDA.

NVIDIA needs to fix it otherwise games just won't use it.

Agreed. After all Just Cause 2 is only game I could think of that used CUDA.

 

[Hiryougan]

Great, async compute for Deus Ex. Which means it'll run like shit on GTX 900 cards. Thanks NVIDIA for your "complete" DX12 support.

I don't think you understand how AC works.
For example if it wasn't used at all in AoS, the results for nvidia would be EXACTLY THE SAME, not better. It simply helps the performance on GCN cards. It doesn't make cards that are not able to do it perform worse.

 

[applejack]

It is doable, Just Cause 2 had a water simulation done in pure CUDA.

Agreed. After all Just Cause 2 is only game I could think of that used CUDA.

if by "pure CUDA" you mean game devs leverage CUDA directly (no middle-ware) - add these:
CUDA/GPU Transcode used in Rage & Wolfenstein (maybe available to other ID Tech 5 titles by ini tweak)
Nascar '14 use CUDA to accelerate in-house particle effects.

however other ~45 games are using CUDA for advanced PhysX acceleration. a middle-ware designed by the creators of the architecture itself is "pure" enough for me.

 

[FordGT90Concept]

I don't think you understand how AC works.
For example if it wasn't used at all in AoS, the results for nvidia would be EXACTLY THE SAME, not better. It simply helps the performance on GCN cards. It doesn't make cards that are not able to do it perform worse.

Async on causes NVIDIA cards to lose performance across the board: http://www.anandtech.com/show/10067/ashes-of-the-singularity-revisited-beta/6

Developers need to enable async compute on GCN cards and disable it on NVIDIA cards to get the best framerate for their respective platforms.

 

[RejZoR]

It's the same situation as with tessellation. If you use it to gain performance, you'd have gains on better hardware and no penalty on "unsupported" hardware. But when you use a feature t cram more "details" into a game, that simply isn't true anymore. And the same is with async compute. It's not there solely to boost identical graphics quality on all graphic cards, I bet they'll use it to cram more details into the game thanks to those gains and in that case, performance just wont' be the same.

What I'm saying is that they won't be using async compute to achieve insane framerate across the board, they'll use it to make more details and sacrifice performance with it. It has always been like this. Instead of making current game more adoptable by more players they want to make it more appealing to the rich crowd with beefed up PC's. And then they wonder why sales aren't up there. Lol...

 

[Hiryougan]

Async on causes NVIDIA cards to lose performance across the board: http://www.anandtech.com/show/10067/ashes-of-the-singularity-revisited-beta/6

Developers need to enable async compute on GCN cards and disable it on NVIDIA cards to get the best framerate for their respective platforms.

Afaik Nvidia cards lost some performance there because they tried to emulate ac(and we see how that went).

 

[BiggieShady]

if by "pure CUDA" you mean game devs leverage CUDA directly (no middle-ware) - add these:
CUDA/GPU Transcode used in Rage & Wolfenstein (maybe available to other ID Tech 5 titles by ini tweak)
Nascar '14 use CUDA to accelerate in-house particle effects.

however other ~45 games are using CUDA for advanced PhysX acceleration. a middle-ware designed by the creators of the architecture itself is "pure" enough for me.

I meant "pure" CUDA in a sense that one part of rendering pipeline (e.g. simulated water geometry in JC2) that would normally be done via geometry shader is done via shared side buffers using CUDA only.
Incidentally that code in JC2 was written by nvidia themselves and probably most of the cuda specific code from your examples and half of the code from 45 instances of middleware integration. The other half came from SDK examples also written by them. So, yes, it's also "pure" that way too. Untainted by all the different middleware modifications or independent CUDA+DirectX engine implementations throughout the world.

 

[FordGT90Concept]

Afaik Nvidia cards lost some performance there because they tried to emulate ac(and we see how that went).

They did not. Async shaders is part of the DirectX 11 API. AMD cards handles async workloads asynchronously where NVIDIA cards handles async workloads synchronously; hence, performance boost for the former and performance hit for the latter. If those 2-4% frames are important, the only solution on NVIDIA is to not use async shaders at all.

 

[BiggieShady]

NVIDIA cards handles async workloads synchronously ... the only solution on NVIDIA is to not use async shaders at all.

That way too oversimplified. Both architectures benefit from async shaders and they will be used simply because it's in direct x. Difference is GCN is more efficient when 3d and compute commands are mixed, which is worst case for nvidia. Nvidia arch likes them separated (in lowest number of batches possible) which is worst case for AMD.
All serious engines will commit workload for a single frame differently on different gpu architectures to achieve peak efficiency.
GCN is great because it allows more simple/flexible approach where you can simply keep brute force saturating command queues. With nvidia currently you gotta treat async task invokes like draw calls, manage and batch them because of the time overhead from context switch. Not only time overhead because synchronicity here is side effect from context switch that may wait on longer running async tasks to finish.