AMD: R8xx Speculation

[rpg.314]

It is likely that this gen or next gen gpu's will ditch the minimum 2x2 quad pixel shader runs, even for a pixel sized triangles to save useless pixel shader work.

 

[mczak]

It is likely that this gen or next gen gpu's will ditch the minimum 2x2 quad pixel shader runs, even for a pixel sized triangles to save useless pixel shader work.

Given that simd width is much larger than that 2x2 quad anyway seems highly unlikely.

 

[DieH@rd]

Look what I did 6 years ago: http://users.belgacom.net/xvox/
Hardware only displacement mapping, even with trilinear filtering of the displacement maps.

If you press the A key, the continuous level of detail tessellation will be adapted so that the animation runs at 60 Hz on any card !

In wireframe mode you can see the tessellation pattern, slightly nicer than what ATI does if I may say.

Excelent demo.

 

[Silent_Buddha]

I'm still wondering if they did anything new to boost the maximum triangle rate of small triangles. Tessellating to pixel sized triangles is fine, but if you can do only like 500 million triangles a second, you then also can do only about 500 million rendered pixels per second...
Shaders and backend must be running at less than 10% efficiency in that case.

Well one of the main points for Tesselation was to increase the number of polygons that can be rendered per scene, so I can only imagine the maximum triangle amount will go up, obviously only when tesselating however.

Early tesselation demo's showed a quite radical increase in triangle count without a large performance hit up to a point. Unfortunately I don't think any of those demo's were made publicly available.

Regards,
SB

 

[Voxilla]

Well one of the main points for Tesselation was to increase the number of polygons that can be rendered per scene, so I can only imagine the maximum triangle amount will go up, obviously only when tesselating however.

One of the Siggraph papers elaborated on how the HW rendering pipeline would need to be modified to efficiently render loads of small triangles. There is quite a bit of work needed.

In my opinion it is not very likely that the R870 already would do so.

 

[Acert93]

While a lot of small triangles may not be efficient I wonder at what point you take that route only because it offers the most detail per pixel. Real geometry via displacement maps, even if a big performance hit, looks really nice compared to POM or normals, and if you are running at nearly 100fps anyhow why not take the hit for some real geometry?

 

[Voxilla]

While a lot of small triangles may not be efficient I wonder at what point you take that route only because it offers the most detail per pixel. Real geometry via displacement maps, even if a big performance hit, looks really nice compared to POM or normals, and if you are running at nearly 100fps anyhow why not take the hit for some real geometry?

Sure, detailed real geometry is best, and with higher resolution screens the more you need of it. It hurts to see that even new games like Rage show angular curved objects.

 

[AnarchX]

5870: 157 fps @ Vantage perlin noise

 

[Lightman]

5870: 157 fps @ Vantage perlin noise

Worth noting is that HD5870 scores were obtained on High preset! Makes it even more impressive :smile:

 

[AnarchX]

Feature tests are afaik not influenced by preset.

 

[fellix]

Yup -- nice guy. I requested results from the two fill-rate tests from 3DMark'06, to see whether the ROPs are BW bottlenecked during alpha-blending.

The cloth simulation feature test in Vantage is still suspiciously slow even on the latest ATi HW, at least compared to ye oldie G80.

 

[rpg.314]

As we discussed earlier, the explosion of geometry thanks to tesselation will likely lead to more emphasis on deferred shading. A possible win for larrabee here.

 

[Humus]

It is likely that this gen or next gen gpu's will ditch the minimum 2x2 quad pixel shader runs, even for a pixel sized triangles to save useless pixel shader work.

No. It needs 2x2 to be able to compute a gradient, needed for any regular mipmapped texture lookup.

 

[MfA]

It's a little more work to really calculate perspective correct differentials for texture coordinates per pixel instead of the difference between pixels, but it's by no means impossible.

 

[snarfbot]

I'm still wondering if they did anything new to boost the maximum triangle rate of small triangles. Tessellating to pixel sized triangles is fine, but if you can do only like 500 million triangles a second, you then also can do only about 500 million rendered pixels per second...
Shaders and backend must be running at less than 10% efficiency in that case.

i remember reading someplace that the dx11 geforce wont have a dedicated tesselator and will be doing all the work in the shaders, maybe thats why.

 

[Ailuros]

i remember reading someplace that the dx11 geforce wont have a dedicated tesselator and will be doing all the work in the shaders, maybe thats why.

Houston I lost connection....

 

[trinibwoy]

Houston I lost connection....

Why? I really haven't seen any documentation on how heavy tessellation would be on the ALUs. I thought the fixed function stuff was just to standardize the behavior of that stage. Doesn't mean it'll necessarily run slowly on programmable hardware.

 

[Ailuros]

Why? I really haven't seen any documentation on how heavy tessellation would be on the ALUs. I thought the fixed function stuff was just to standardize the behavior of that stage. Doesn't mean it'll necessarily run slowly on programmable hardware.

That's one reason why I lost the connection; the second is that I can't figure out what the tesselation unit (or lack thereof) has directly to do with small triangle efficiency.

 

[CouldntResist]

That's one reason why I lost the connection; the second is that I can't figure out what the tesselation unit (or lack thereof) has directly to do with small triangle efficiency.

Perhaps it's not worth to have dedicated unit, when it's desired usage leads to inevitable inefficiencies in other areas of hardware...

Tesselation == checkbox feature, to satisfy DX certificate?

 

[ECH]

Yup -- nice guy. I requested results from the two fill-rate tests from 3DMark'06, to see whether the ROPs are BW bottlenecked during alpha-blending.

The cloth simulation feature test in Vantage is still suspiciously slow even on the latest ATi HW, at least compared to ye oldie G80.

Express what you really want to say about that