AMD confirms GCN, incl. XBO, doesnt support DX12 12_1

BTW guys!! is witcher 3 gonna get a directx 12 patch after windows 10 launch..

 

GTA V needs a DX12 version badly.

 

http://en.wikipedia.org/wiki/Direct3D#Direct3D_12_levels acording to this nvidia made up there own dx12 to bypass tier 3 since they are only tier 2 cards but amd is fully tier 3 and have no limitations unlike the nvidia cards

 

As an extreme layman, with a R9 280X and overclocked 2500K, can someone explain this to me in stupid speak what this will or won't mean for me?

 

AMD didn't remove the support for DX12. Microsoft has changed something in the latest builds of Windows 10 and hence, it's not working at the moment.

I've just tested using the same drivers in the current build 10130. It doesn't work.

I hope that in newer builds of Windows 10 and drivers, we can test it again.

 

Terascale 2 was running Direct X 11 feature level 10_1. GCN jumped straight to Direct X 11 Feature level 11_1. I had forgot about that little detail.

But yeah, on AMD's side only GCN will support DX12.

 

Thanks. I will re-phrase that for next time

 

Indeed there seemed like a change in this latest 10130 build. Let us see on what will happen.

 

The same it meant for the 780Ti not to have DX 11.2

Nothing.

 

You are completely right. We have feature "levels" and DX12 tiers. Maxwell v2 (900 series) will be Tier 2. All GCN cards will be Tier 3. Even old cards like HD7770.

So OP is really wrong with "go buy nv" if you want full DX12 support thing. He should really remove it, we have enough false nv marketing already.
And looks like there will be no DX12 support for HD5000/6000 series. But there will be some (Tier 1) for Fermi cards (GTX400/500).

 

edit: double post :S

 

Truth is that enough FUD has been spread around lately, in here too. It is truly a pity for the 5000/6000 series, but the fixed architecture didn't leave much to do. Fermi on the other hand was a much more forward-thinking design at the time (remember Linux running JUST on the GPU? That was Fermi), and that paid off, hence DX12 support for GTX 4xx cards and on.

The inner geek in me is happy that the new Radeon card seems to be hot, since it means that once more AMD didn't compromise on the compute power and general purpose nature of the architecture.

 

Not "DX12" tiers, but five separate optional capabilities which have "tiers": resource binding tiers, tiled resource tiers, conservative rasterization tiers, resource heap tiers, and cross-node sharing tiers.

Feature level 12_0 is supported by GCN 1.1 Bonnaire (R7 260, R9 M280, HD 7790/8770), Hawaii (R9 290), and GCN 1.2 Tonga (R9 285, R9 M295), Iceland (R5 M240/M250/M255 from end of 2014), and Fiji (Radeon Fury).

And BTW it's time to stop this "fully compatible" nonsense. We have an new API that offers like [post="5090538"]several thousand baseline features[/post] vital for getting the most out of current advanced graphics hardware that rivals supercomputers of the past. Yet everyone has to whine about 3 (three) minor features that no-one will be using any time soon... beats me completely.

All GCN and Xbox One are Resource binding Tier 3. Fermi and Haswell/Broadwell are Resource binding Tier 1, everything else is Tier 2 (Skylake, Kepler, Maxwell-1 and Maxwell-2). Each higher tier is a superset of a lower.

RB Tier 1 GPUs make 39% of all Direct3D 12 compatible hardware on the desktop, Tier 2 is 44.3% and Tier 3 is 16.7% according to the results of February 2015 Steam hardware survey.


en.wikipedia.org/wiki/Direct3D#Feature_levels

GCN 1.1/1.2 support feature level 12_0 and GCN 1.0 supports feature level 11_1 - but it also supports two of the three mandatory features of level 12_0 with the exception of Tiled Resources Tier 2, [post="5089840"]which is really a very minor improvement over Tier 1[/post] that can be easily worked around.

And of course the feature levels are nested in strict supersets, so each higher level includes features from all lower levels.


As to resource binding tiers, the Direct3D 12 API is designed to be bindless - everything is in resource descriptor heaps/tables, and there is no automatic resource/memory management on the API side. You just populate the heaps with links to resources (shaders, textures, geometry data etc.) loaded in local video or system memory.

However older hardware imposes restrictions on both the total number of descriptors for so-called "access views" (UAVs, CBVs, SRVs) - which allow random read/write access to the specified memory resource - and the number of these decriptors and texture samplers that can be passed to each shader stage. So for Resource Binding Tier 1 (not bindless) or Tier 2 (partially bindless) hardware, developers must explicitly track these limits in their applications, but on Tier 3 (fully bindless) hardware, there are no practical limits.


Bindless tiers allow huge shaders with lots of objects and unlimited textures and samplers, etc. and will also benefit ExecuteIndirect and MultiGPU which use built-in Async Copy / Execute / Compute engines (a base Direct3D 12 feature) to allow any GPU and CPU access any resource in any system or local memory. This is a huge relief for programmers trying to implement advanced rendering algorythms and efficient CPU/GPU or Multi-GPU synchronisation.

These are not "instruction sets", HLSL remains at Shader Model 5.x

Feature level 12_1 includes two new pipeline states for the fixed function rasterization hardware, rasterizer-ordered pixel processing and fine-tuned rasterisation algorythm, which are currently only supported by Maxwell-2.

Feature level 12_0 exposes "virtual memory" capabilities (dynamic heaps for resource descriptor tables, UAV descriptor tables for typed texture formats, virtual memory "paging" for Texture2D resources) of the GPU memory management unit which were not supported in earlier versions of DXGI/WDDM.

None of these affect performance in any significant way (with the exception of ROVs, which are very efficient for order-independent transparency).

Again, the only feature that differentiates GCN 1.0 parts conforming to level 11_1 from newer GCN 1.1/1.2 parts with level 12_0 is only tiled resources tier.

Here is [post=5089840]the exact difference between tiled resource tiers 1 and 2[/post]:

a) if a virtual memory page is not present in GPU memory, tier 1 hardware returns garbage on reads and raises an exception on writes, so the application has to check page residency first; tier 2 returns zero on reads and silently discards any writes;
b) you can use LOD level clamp with tiled textures on tier 2 and read back tile mapping status in the shader code; on tier 1, you have to emulate LOD level clamp with shader code and uwe the AP call to check the statusI;
c) MIP levels spanning multiple tiles are guaranteed to not use hardware-specific packing formats on Tier 1 if they are an integer multiple of a tile shape, while on Tier 2 this expands to MIPs that can fit in a single tile.

Now show me one single game that uses tiled resources.

Doh.

Evergreen (HD 5000) has always been at feature level 11_0.

Direct3D 12 works fine with build 10130 and driver 1023.x

 

What do you mean by "Direct3D 12 works fine"?

Because as far as the API Overhead Test of 3DMark, DX12 hasn't worked since build 10074 or so for any of us around here.

 

Bookmarked. Thanks for this post, it is very clear and informative

 

Tier 3 is the best tier fact is maxwell cards don't fully support dx12 they are limited in a few ways but amd GCN isnt

http://en.wikipedia.org/wiki/Direct3D#Direct3D_12_levels

Direct3D 12 requires graphics hardware conforming to feature levels 11_0 and 11_1 which support virtual memory address translations. It introduces a revamped resource binding model, allowing explicit control of memory using descriptor heaps and tables. This model is supported on majority of existing desktop GPU architectures and requires WDDM 2.0 drivers. Supported hardware is divided into three Resource Binding tiers, which define maximum numbers for descriptor heaps used for CBV (constant buffer view), SRV (shader resource view) and UAV (unordered access view); CBVs and SRVs per pipeline stage; UAVs for all pipeline stages; samplers per stage; and SRV descriptor tables. Tier 3 hardware such as AMD GCN has no limitations, allowing fully bindless resources managed through dynamic memory heap, while Tier 1 (Nvidia Fermi, Intel Haswell/Broadwell) and Tier 2 (Nvidia Kepler/Maxwell, Intel Skylake) hardware impose some limits on the number of these resources

 

I've successfully run the 3DMark API Overhead feature test (including DX12) thanks to Windows 10 build 10134 and Catalyst v15.200.1023.5:

Spoiler

Full test results:
http://www.3dmark.com/3dm/7227092?

Cheers!

 

I though amd_roy was not in this DX12 "PR party" but i was wrong.

This article explains AMD and Nvidia DX12 support levels per GPU and has some simple graphical examples of different DX12 features.

Text link to article:

http://pastebin.com/F5qbYpA9

What i fear the most in it was the ABSOLUTE SECURITY amd_roy has about DX12 support from AMD:

 

He's correct man. Look at DX11.2. For some things AMD GPU's even have MORE support than NVIDIA cards, I don't see anybody from NVIDIA fretting. As it was posted above with links and everything, both manufacturers support something that the other does not, among the thousands of functions they both have in common.

 

I can trust in your comments, i can't say the same for "next week" man...