GCN can't support Conservative Rasterization or Rasterizer Ordered Views

[kagui]

well accordingly to pcper on 3dmark (dx12 version ), gcn is getting a big boost, the 290x tides with a titan x, on draw calls only, so the performance that still is missing from the equation is the shader, but that cant be a much of a diff.

 

[omek]

Not really note-worthy...
Even NVIDIA says that it can be achieved without hardware support through shaders because it's essentially just an algorithm (which means that GCN, Fermi and Kepler could support it).
https://developer.nvidia.com/content/dont-be-conservative-conservative-rasterization
https://developer.nvidia.com/gpugems/GPUGems2/gpugems2_chapter42.html

 

[Silverforce11]

The DX12 slide from brazilian source. Is it fake? Or what?

Asynchronous DMA & Compute are features that enhance rendering & compute efficiency. This was to be in the next stage of Mantle which appears to be dead. Glad to see it make its way into DX12!

Also, its clear Kepler lacks *many* DX12 features and Maxwell v2 makes up for it, somewhat.

So comparing GCN to Maxwell v2, its trading CR & ROV for quite a few other features which Maxwell v2 lacks! Not bad for such an old architecture.

 

[Abwx]

So GCN doesn't support DirectX 12 to the full extent that Maxwell does.

That s the other way around...

The DX12 slide from brazilian source. Is it fake? Or what?

It is right, only GCN cards support the tier_3, Nvidia cards are limited to tier_2 and do not support all functionalities of DX12 even if they are branded 12_1, indeed this later level has been tailored by MSoft for Nvidia, and not because they support DX12 fully, guess why...

http://www.hardware.fr/html/news/cat4/page4.html#actu_14135

 

[nvgpu]

Resource Binding Tier is irrelevant, only Feature Levels are relevant. Also that Resource Binding Tier image is OUTDATED & WRONG now, but keep posting it I guess.

https://forum.beyond3d.com/threads/direct3d-feature-levels-discussion.56575/page-5#post-1834185

Hardware tiers for resource binding/descriptor heaps:
https://msdn.microsoft.com/en-us/library/windows/desktop/dn899127(v=vs.85).aspx

Maximum UAVs is now 64 on Tier 1 and unlimited on higher Tiers, so the tables are not valid anymore.

 

[garagisti]

Resource Binding Tier is irrelevant, only Feature Levels are relevant. Also that Resource Binding Tier image is OUTDATED & WRONG now, but keep posting it I guess.

https://forum.beyond3d.com/threads/direct3d-feature-levels-discussion.56575/page-5#post-1834185

Right. Cause it is nothing like what happened with Vista and DX10 and being repeated by Nvidia and MS, once again.

 

[Silverforce11]

Now that we are 99.9% sure that DX12 is built upon Mantle as the foundation (hands up if you are still in denial!), it makes sense to see that GCN actually support all of its features, except for FL12.1 added by NV for Maxwell v2 architecture.

To expect GCN to be "less compatible" with DX12 is silly when the API is built around what GCN is capable of.

 

[nvgpu]

GCN 1.0 can't support Feature Level 12_0, this has been proven already since it only supports Tiled Resources Tier 1 as reported by DX Caps Viewer and Tier 2 is the minimum requirement set by MS. So claims like "GCN actually support all of its features" is already wrong.

 

[ShintaiDK]

The current state as MS sees it. Anything else is irrelevant.

 

[Fallen Kell]

You accidentally sneaky missed something when quoting Ryan

(These are beta drivers on a not even beta OS, so as usual the standard disclaimers about results being subject to change do apply. Especially if not all features are currently being exposed in these drivers)

PS: I'm not saying that it's not true, just that it's an important point.

I wouldn't consider the Win10 build that is out a "not even beta OS" given the fact that MS is releasing in just 3-4 months.

 

[Abwx]

GCN 1.0 can't support Feature Level 12_0, this has been proven already since it only supports Tiled Resources Tier 1 as reported by DX Caps Viewer and Tier 2 is the minimum requirement set by MS. So claims like "GCN actually support all of its features" is already wrong.

You make it sound like AMD s cards are less advanced while that s because of Nvidia s uarch limited flexibility, MS had to create a level feature specificaly for Nvidia, because these are not versatile uarchs..

Here HFR word about it, you can check if my translation is accurate with google :

Selon nos informations, les GPU Maxwell sont en fait limités au Tier 2, qui est nécessaire au support des niveaux 12_0 et 12_1 et qui a probablement lui aussi été modelé autour de leurs capacités. Une des différences les plus importantes avec le Tier 3 concerne la gestion des Constant Buffer Views (CBV) : ceux-ci ne sont pas virtualisés et sont limités en nombre à 14. Il est probable que l'architecture Maxwell soit capable de virtualiser les CBV, mais que l'implémentation logicielle/matérielle de Nvidia profite d'un mode plus performant avec une gestion "fixe" des Constant Buffers. Un compromis qui limite quelque peu la flexibilité accordée aux développeurs pour s'assurer que les GPU GeForce restent dans un mode optimal sur le plan des performances.

My translation :

According to our informations Maxwell GPUs are actualy limited to Tier 2, wich is necessary for 12_0 and 12_1 levels and wich has been probably modeled around their capabilities.

One of the most important difference is about management of Constant Buffer Views (CBV) : theses are not virtualised and are limited to 14.
It is likely that Maxwell uarch is capable to virtualise the CVBs, but that Nvidia s software/hardware implementation benefit from a more performing mode with a "fixed" management of Constant Buffers.
A compromise that somewhat limit the flexibility awarded to developpers to ensure that the Geforce GPUs stay within an optimal mode for performances.

http://www.hardware.fr/html/news/cat4/page4.html#actu_14135

And the journalist is a known Nvidia fan...

 

[Noctifer616]

Asynchronous DMA & Compute are features that enhance rendering & compute efficiency. This was to be in the next stage of Mantle which appears to be dead. Glad to see it make its way into DX12!

I seen Asyn DMA and Compute in Mantle slides, was talked about back at GDC 2014. Both the Thief and BF4 guys said that Mantle supports Asyn Compute, but neither game used it.

Also, I have seen a screenshot from a test running Asyn Compute, not sure if it was Mantle though.

Edit: the demo was in DirectX 12, not Mantle.

Edit 2: The Mantle documentation also mentions support for Asyn DMA and Compute.

Mantle API defines two queue types: a universal queue (GR_QUEUE_UNIVERSAL) and an asynchronous compute queue (GR_QUEUE_COMPUTE). Other queue types, such as DMA and so on are exposed through extensions. There is at least one universal queue available for the Mantle device; other queues are optional.

 

[Spanners]

Where is Thatbuzzkiller now? And I was the troll for him..

Not sure why you'd be looking for him. You contributed to getting this interesting thread locked with your repeated insistence that GCN didn't support tier 3 Resource Binding model. Never saw him say that GCN had hardware support for Conservative Rasterization or Rasterizer Ordered Views which is this threads topic. So why are you here implying he's gone missing because the information in this thread proves him wrong?

I may have missed something I don't read every thread but it seems you were the one that was incorrect.

 

[Shehriazad]

All this DX12 madness makes me...mad.

DX12 ain't even out...DX12.1

"All DX11 cards can run DX12 stuff"

Well...apparently that is going to be so limited that those cards will drop right out as soon DX12 games actually show up.

Like damn...Microsoft be Microsofting?


That said....I won't really have need for DX12 until mid/end 2016, anyway....and by then the 2nd gen HBM cards are out...so I'm totally cool with it.

 

[waffleironhead]

https://forum.beyond3d.com/threads/direct3d-feature-levels-discussion.56575/page-4#post-1833996



That settles the matter and DX Caps Viewer is always right. DX Caps Viewer vindicated what I said, GCN can't support Direct3D12 fully.

Maxwell2 is Tier 3 Tiled Resources(GCN1.1/1.2 Tier 2 only and GCN 1.0 Tier 1 only) and fully supports Conservative Rasterization & Rasterizer Ordered Views.

Well I wouldn't feel too smug if I was you. You claimed Maxwell supported both and according to the linked image you left out of Ryan's quote it doesn't. So 1 right out of 2 ain't bad right?

 

[ThatBuzzkiller]

Where is Thatbuzzkiller now? And I was the troll for him..

How real mature of you ...

There's some incorrectly listed information in this picture ...

For instance, first gen GCN already has support for Tiled Resources Tier 2, not Tier 1. The title of this thread along with a part of this table regarding GCN supporting ROVs is incorrect since it supports the OpenGL equivalent of this feature. All AMD needs to do is update their drivers to get support for ROVs.

Furthermore it is likely that GCN supports Tier 2 of Typed UAV loads since the architecture supports many more types of format for type conversion reads.

The only known architectures to support Pixel Shader Specified Stencil Reference Value so far is AMD R600 and up! Neither Nvidia or Intel makes an effort to support it or get a similar OpenGL extension!

It's clear that 2nd Maxwell doesn't support EVERYTHING that DX12 has to offer so I'm not exactly why AMD is getting barked at mostly for not supporting everything that DX12 offers when it's the same for Nvidia!

 

[Silverforce11]

Yup, it's interesting that what started with dissing on GCN is actually revealing far more regarding the lack of DX12 features for Kepler, Maxwell v1 and even v2!

 

[JDG1980]

The current state as MS sees it. Anything else is irrelevant.

Interesting chart. I wasn't aware that there were that many differences between GM1xx (GTX 750/750Ti) and GM2xx (GTX 900 series). So the rundown is as follows:

DirectX 12_1: GM2xx only
DirectX 12_0: GCN 1.1 and 1.2
DirectX 11_1: GCN 1.0
DirectX 11_0: GM1xx, Kepler, and Fermi

AMD can therefore claim to support "DirectX 12" for Bonaire, Hawaii, Tonga, and presumably any upcoming releases. This chart is one more reason why I don't believe the rumors that AMD is going to rebrand Pitcairn yet again: they're not going to want to have a major midrange part in their lineup completely lack DirectX 12 support. It doesn't even support AMD's own Mantle API fully (presumably for the same hardware reasons).

If Nvidia eventually releases a GTX 950, I wonder if it will just be a rebrand of the existing GM107, or if they'll update it to a new GM207 chip instead. I suspect the latter, since Nvidia seems to be fairly willing to respin their silicon.

 

[Spanners]

"DirectX 12_1: GM2xx only
DirectX 12_0: GCN 1.1 and 1.2
DirectX 11_1: GCN 1.0
DirectX 11_0: GM1xx, Kepler, and Fermi"

These are just feature levels inside DX12 the absolute minimum for "support" is resource binding Tier 1 + full DX11 support. So Pitcairn doesn't completely lack DX12 support.

 

[pTmdfx]

I seen Asyn DMA and Compute in Mantle slides, was talked about back at GDC 2014. Both the Thief and BF4 guys said that Mantle supports Asyn Compute, but neither game used it.

Also, I have seen a screenshot from a test running Asyn Compute, not sure if it was Mantle though.

Edit: the demo was in DirectX 12, not Mantle.

Edit 2: The Mantle documentation also mentions support for Asyn DMA and Compute.

At least for Battlefield 4, it is used, or more precisely, Frosbite makes use of it. The engine runs quite a large set of compute shaders in parallel to the main rendering pipeline.

 

[ViRGE]

Interesting chart. I wasn't aware that there were that many differences between GM1xx (GTX 750/750Ti) and GM2xx (GTX 900 series). So the rundown is as follows:

DirectX 12_1: GM2xx only
DirectX 12_0: GCN 1.1 and 1.2
DirectX 11_1: GCN 1.0
DirectX 11_0: GM1xx, Kepler, and Fermi

AMD can therefore claim to support "DirectX 12" for Bonaire, Hawaii, Tonga, and presumably any upcoming releases. This chart is one more reason why I don't believe the rumors that AMD is going to rebrand Pitcairn yet again: they're not going to want to have a major midrange part in their lineup completely lack DirectX 12 support. It doesn't even support AMD's own Mantle API fully (presumably for the same hardware reasons).

If Nvidia eventually releases a GTX 950, I wonder if it will just be a rebrand of the existing GM107, or if they'll update it to a new GM207 chip instead. I suspect the latter, since Nvidia seems to be fairly willing to respin their silicon.

Arg! It's not "DirectX", it's "Feature Level"!

All of these GPUs support DirectX 12. Full stop. What you are trying to argue is the feature levels, of which DirectX 12 includes many.

Please for the love of the sanity of everyone who actually knows how this stuff works, stop calling the feature levels DirectX.

 

[StereoPixel]

AMD said that all GCN supports Tiled Resources Tier2 as well

 

[therealnickdanger]

Not really note-worthy...
Even NVIDIA says that it can be achieved without hardware support through shaders because it's essentially just an algorithm (which means that GCN, Fermi and Kepler could support it).
https://developer.nvidia.com/content/dont-be-conservative-conservative-rasterization
https://developer.nvidia.com/gpugems/GPUGems2/gpugems2_chapter42.html

That's what I understood as well. Even if a particular DX12 feature level is not natively supported in hardware, AMD/NVIDIA/Intel/Matrox(lulz) can still dedicate GPU resources to process those commands via software. Would it be as potent as native hardware? Probably not. What will the impact be? Unknown.

Until we have a more matured (production) API and drivers tested on actual hardware, fighting over the speculative nature of these data is pointless.

 

[Mindtaker]

Maxwell is feature 12_1.
GCN 11.1 and 1.2 is feature 12_0.

The information is very clear, let it go.

 

[csbin]

Maxwell is feature 12_1.
GCN 11.1 and 1.2 is feature 12_0.

The information is very clear, let it go.

Only Maxwell 2 is feature 12_1

Maxwell 1 is feature 11_0