Fascinating article on DX11

[soccerballtux]

Here.

Very good read.

 

[ViRGE]

Both R6xx and R7xx hardware have tessellators, but since these are proprietary implementations, they won't be directly compatible with DirectX 11 which uses a much more sophisticated setup.

That must suck for AMD.

 

[SunnyD]

Originally posted by: ViRGE

Both R6xx and R7xx hardware have tessellators, but since these are proprietary implementations, they won't be directly compatible with DirectX 11 which uses a much more sophisticated setup.

That must suck for AMD.

Oh I don't know about that. My guess is there will be a DX wrapper call that enables functionality, similar to the way GL Extension work.

And speaking of - I can tell you that the tessellators are a GOOD thing... probably one of the better things put into graphics technology in the past 5 years. I'd rate it far more important than PhysX and Stereoscopic but at least an order of magnitude.

 

[TridenT]

Too bad I don't understand a word...

 

[thilanliyan]

Originally posted by: SunnyD
And speaking of - I can tell you that the tessellators are a GOOD thing... probably one of the better things put into graphics technology in the past 5 years. I'd rate it far more important than PhysX and Stereoscopic but at least an order of magnitude.

Shoot you had to go there. Cue the "debates".

 

[habbakuk87]

Tessellators must be real important as one of the two "real gaming consoles this gen" already has and uses them.

 

[NoStateofMind]

Originally posted by: soccerballtux
Here.

Very good read.

:thumbsup:

 

[AzN]

I didn't read the article but is MS dx11 trying to save bandwidth and fillrate by using tessellation like what ST micro did with Kyro cards? Why are they trying to adopt this now? what ever happened to Kyro?

By looking at the charts it looks like nothing but a marketing gimmick when in fact it's not going to change much.

 

[geokilla]

I gave up reading on that article after page 2. Got too technical for us noobs.

 

[SunnyD]

Originally posted by: Azn
I didn't read the article but is MS dx11 trying to save bandwidth and fillrate by using tessellation like what ST micro did with Kyro cards? Why are they trying to adopt this now? what ever happened to Kyro?

By looking at the charts it looks like nothing but a marketing gimmick when in fact it's not going to change much.

Tessellation != TBR.

Tessellation is where the model is represented only by point vertices, and the tessellator is responsible for converting the point map into a solid geometric model. It is a bandwidth savings, but also a very expensive algorithm. (It's interesting to see the way it's described for DX11 - looks simply like a LOD tool)

Kryo uses something known as "Tile-based Rendering". It breaks down a scene into small tiles, and in a manner similar to something like SLI, allows the video processor to render a scene through multiple rendering pipelines at the same time. Kind of "SSE" for video.

 

[taltamir]

Originally posted by: ViRGE

Both R6xx and R7xx hardware have tessellators, but since these are proprietary implementations, they won't be directly compatible with DirectX 11 which uses a much more sophisticated setup.

That must suck for AMD.

well, I am sure AMD should be able to make revisions to it to make it match the DX11 requirement with less work than nvidia can create one from scratch. And probably make a wrapper that will at least do part of the DX11 tesselation on current AMD hardware.

 

[ViRGE]

Originally posted by: taltamir

Originally posted by: ViRGE

Both R6xx and R7xx hardware have tessellators, but since these are proprietary implementations, they won't be directly compatible with DirectX 11 which uses a much more sophisticated setup.

That must suck for AMD.

well, I am sure AMD should be able to make revisions to it to make it match the DX11 requirement with less work than nvidia can create one from scratch. And probably make a wrapper that will at least do part of the DX11 tesselation on current AMD hardware.

I'm sure they can come out with a DX11 tesselator (I mean really, it's not like they have a choice), but as for a wrapper, I wouldn't bet on one. But in the present, that die space they used to include a tesselator on R6xx and R7xx products has been wasted if it's not going to be DX11 compatible, especially since no one is going to want to program specifically for it in lieu of targeting DX11.

 

[SickBeast]

So will DX11 take advantage of CUDA and such?

 

[taltamir]

no. although nvidia could choose to program some of the functions of DX11 in cuda. although that might be less efficient then coding it to run directly on the hardware without cuda translations.

 

[Denithor]

Nope, talt's right, using CUDA would just add another layer of overhead. DirectX is just another API that will allow the programmers to use the hardware, like CUDA but specifically for graphics acceleration where CUDA is general-purpose.

 

[AzN]

Originally posted by: SunnyD

Originally posted by: Azn
I didn't read the article but is MS dx11 trying to save bandwidth and fillrate by using tessellation like what ST micro did with Kyro cards? Why are they trying to adopt this now? what ever happened to Kyro?

By looking at the charts it looks like nothing but a marketing gimmick when in fact it's not going to change much.

Tessellation != TBR.

Tessellation is where the model is represented only by point vertices, and the tessellator is responsible for converting the point map into a solid geometric model. It is a bandwidth savings, but also a very expensive algorithm. (It's interesting to see the way it's described for DX11 - looks simply like a LOD tool)

Kryo uses something known as "Tile-based Rendering". It breaks down a scene into small tiles, and in a manner similar to something like SLI, allows the video processor to render a scene through multiple rendering pipelines at the same time. Kind of "SSE" for video.

TBR? Tile based Rendering?

It seems MSdx11 and Kyro uses same kind of technology. Dx11 a bit more advanced using shader while Kyro used Rop.

Why is tessellation expensive? LOL...

 

[Creig]

Originally posted by: ViRGE

Originally posted by: taltamir

Originally posted by: ViRGE

Both R6xx and R7xx hardware have tessellators, but since these are proprietary implementations, they won't be directly compatible with DirectX 11 which uses a much more sophisticated setup.

That must suck for AMD.

well, I am sure AMD should be able to make revisions to it to make it match the DX11 requirement with less work than nvidia can create one from scratch. And probably make a wrapper that will at least do part of the DX11 tesselation on current AMD hardware.

I'm sure they can come out with a DX11 tesselator (I mean really, it's not like they have a choice), but as for a wrapper, I wouldn't bet on one. But in the present, that die space they used to include a tesselator on R6xx and R7xx products has been wasted if it's not going to be DX11 compatible, especially since no one is going to want to program specifically for it in lieu of targeting DX11.

Current AMD tesselators might not necessarily be wasted. Although the R6xx and R7xx GPUs may not contain full support for the DX11 spec, they could still be capable of handling some of the calls. Sort of like how the Nvidia GT200 isn't fully DX10.1 compliant, yet they can still perform some of its features.

 

[SunnyD]

Originally posted by: Azn

Originally posted by: SunnyD

Originally posted by: Azn
I didn't read the article but is MS dx11 trying to save bandwidth and fillrate by using tessellation like what ST micro did with Kyro cards? Why are they trying to adopt this now? what ever happened to Kyro?

By looking at the charts it looks like nothing but a marketing gimmick when in fact it's not going to change much.

Tessellation != TBR.

Tessellation is where the model is represented only by point vertices, and the tessellator is responsible for converting the point map into a solid geometric model. It is a bandwidth savings, but also a very expensive algorithm. (It's interesting to see the way it's described for DX11 - looks simply like a LOD tool)

Kryo uses something known as "Tile-based Rendering". It breaks down a scene into small tiles, and in a manner similar to something like SLI, allows the video processor to render a scene through multiple rendering pipelines at the same time. Kind of "SSE" for video.

TBR? Tile based Rendering?

It seems MSdx11 and Kyro uses same kind of technology. Dx11 a bit more advanced using shader while Kyro used Rop.

Why is tessellation expensive? LOL...

Tessellation algorithms are very expensive - there's a lot of math going on behind the scenes to generate each triangle properly (finding nearest neighbors, ensuring no overlap, etc). I worked on a software tessellator once... I abandoned that because it just wasn't feasible with the hardware at the time.

 

[AzN]

Originally posted by: SunnyD
Tessellation algorithms are very expensive - there's a lot of math going on behind the scenes to generate each triangle properly (finding nearest neighbors, ensuring no overlap, etc). I worked on a software tessellator once... I abandoned that because it just wasn't feasible with the hardware at the time.

I bet in software implementation. I'm not too sure how taxing it would be with modern GPU that has tessellator in hardware.

 

[BenSkywalker]

After reading the article I really hope Derek just missed every good thing in DirectX 11, what he covered in the article all seemed like absolute garbage.

CS is nice from a theoretical standpoint in an abstract sense from a flexibility in coding perspective, but it is a nasty bottleneck in a rendering pipeline if utilized, horribly poor- particularly at the stage it would need to be implemented at. As described in the article it will create massive stalls and be a performance hit far greater then SSAA used to be, multi passing anything he mentioned would be far more efficient. This seems to be something Intel muscled in to the spec to waste die space for AMD and nV.

Then we have tesselators.

Tesselators are nice in theory and would have been great in the DX8 era, but now?

So we use a displacement map and generate tesselation based off of it- all we are getting is additional geometric data for elements which will have no bones tied to them and no sort of IK at all capable of any part outside of that which was part of the initial vertex data. The only thing it can amount to is a souped up version of bump mapping as it is implemented. When we had fixed function hardware and had vertex shaders sitting around idle it would have made sense using them for tesselation as it is better to get something out of them then nothing, but to sacrifice die space for a very primitive addition? To gain what? Some bandwidth savings on vertex data? Bump your RAM up another 5MHZ to compensate for it

I worked on a software tessellator once...

Something along the lines of voxel/NURBS or a vertex based tesselator? I could understand the shift if it we were changing primitives, but as it stands now I just don't see the logic behind it.

 

[SunnyD]

Originally posted by: Azn

Originally posted by: SunnyD
Tessellation algorithms are very expensive - there's a lot of math going on behind the scenes to generate each triangle properly (finding nearest neighbors, ensuring no overlap, etc). I worked on a software tessellator once... I abandoned that because it just wasn't feasible with the hardware at the time.

I bet in software implementation. I'm not too sure how taxing it would be with modern GPU that has tessellator in hardware.

Well of course, dedicated transistors for a fixed function task will definitely be faster! But the algorithms themselves still are pretty intensive stuff.

Originally posted by: BenSkywalker
Something along the lines of voxel/NURBS or a vertex based tesselator? I could understand the shift if it we were changing primitives, but as it stands now I just don't see the logic behind it.

No, 3D triangle setup, primarily for adaptive LOD. Tessellators are great for that sort of thing (partly why DX11 hypes them so much). I was working on GeForce 1/2 class hardware at the time with Pentium II's... and the amount of triangles I was pushing (industrial app) was painful for the hardware of that time. Adaptive LOD would have been a godsend for that particular application.

 

[chizow]

Originally posted by: BenSkywalker
After reading the article I really hope Derek just missed every good thing in DirectX 11, what he covered in the article all seemed like absolute garbage.

Yeah, I didn't really see anything fascinating in there either, the cliffs are basically:

1) Vista sucks, so DX10 adoption sucked too.
2) Win7 is great, so we'll finally see the benefits of DX10 features when DX11 rolls out.

Other than that, the only major hardware requirement added to DX11 was for fixed function tesselator units, everything else is just adding features that are still compatible with DX10 hardware.