I saw this on OcUK today:
http://forums.overclockers.co.uk/sho...php?t=18145991
If true this means the shaders can also be used to assist the tessellator in tessellation heavy scenes.
I saw this on OcUK today:
http://forums.overclockers.co.uk/sho...php?t=18145991
If true this means the shaders can also be used to assist the tessellator in tessellation heavy scenes.
*sigh* a triumph for benchmarks over real world performance.
The reason you don't want to use general purpose shaders for tessellation is that they should be busy doing all the other important things. Only in an artificial situation where they are being under-used should you really gain an advantage from using them for tessellation - a bit like physX really. But I guess AMD have been forced to follow nVidia's benchmark-centric approach![]()
I think this is actually a great idea from ATI and I look forward to it.
Their current HD5XXX series tesselation solution is hardware bottlenecked on anything over "moderate" tesselation, and if you need proof run Unigine or Stone Giant, turn on the wireframe and see what happens when you get the number of triangles up to high!
Nvidia's GTX 4XX solution scales up to crazy levels of geometry without the steep performance drop off curve, (theoretically this is because they beefed up the vertex fetching to 8 per clock). This is a good thing, and undoubtedly superior to ATI at the moment.
ATI have no choice but to focus on the synthetic benchmarks since the only significant tesselating game around at the moment is metro 2033. It's true that some synthetic benchmarks are fairly unrelated to real-world gaming, but for vanilla DX11 tesselation, these benchmarks should be fairly representative.
It's true that the shaders are not "supposed" to be doing tesselation from a purist hardware resolve standpoint, but why not have them help out? All grist to the mill - and shaders these days are extremely general purpose processors!
The trade off from the loss of shaders working on traditional activities, which are now given over to tesselation activites for the same frame, should hopefully equate to better overall efficiency and better minimum frame rates!
We'll see.
I agree with you that the Unigine demo takes the tesselation a bit far, especially on any setting higher than what they call "moderate". Their idea of "normal" tesselation is too strong for my HD5870 and it crumples on the min-framerate with this setting applied and as for "high", forget it.
However, the better example is the Stone Giant demo. This is good in that you can change the tesselation setting inside the demo, and using the wireframe option, look at the number of primitives generated and make up your own mind about the level of geometry on the different models, at a given draw distance.
To me, the stone giant demo looks reasonable in what it terms "medium" and "high" tesselation given the amount of geometry generated.
I take the physx comparison on board, where the shaders are forced to do both physics and graphics rendering simultaneously, giving a possible worst of both worlds scenario in some cases. At least with physx you can ring-fence another GPU (or hopelessly try and use the CPU) - ATI shader tesselation acceleration will only fit the first of these scenarios.
There are currently 1 users browsing this thread. (0 members and 1 guests)