There was a site that had split the benchmarks of various games into their respective DX level, DX9, DX10 and DX11 games, the results showed the 5000 series hardware doing better than Nvidias 400 series in DX9, being pretty much equal in DX10 games and generally losing in DX11 titles, but for the life of me i can't remember what site it was.Originally Posted by Neeyik
They have fixed the dodgy tessalation performance in the 6000 series though right?There was a site that had split the benchmarks of various games into their respective DX level, DX9, DX10 and DX11 games, the results showed the 5000 series hardware doing better than Nvidias 400 series in DX9, being pretty much equal in DX10 games and generally losing in DX11 titles, but for the life of me i can't remember what site it was.
The 5000 series was terrible at DX11.
not `DX11` only really heavy tessallation , which is only a part of DX11
And what are the other "parts"? Direct2D, heterogenous gfx support, GPCPU, some multicore cpu support tweaks and that's about it. Most of the DX11 stuff can be done in DX10 except for tesselation (the relevant part)i believe and that is where AMD sadly gets crushed.
Arguably, tessellation has been a feature of D3D since v8 and it's only the level of control over the algorithms that's anything noticeable with D3D11; in other words, I disagree that it's a key feature compared to the introduction of the Compute Shader. Marketing/PR people would disagree with me simply because it's easier to promote tessellation ("look at all these bumps!!!1"") than the use of dynamic shader linkage ("those big shaders in games....well, they're now...umm, the same... but easier to handle!!!!1!"). We've seen no benchmark that really use SM5.0 yet, with all the new features and limits that it brings; until one such test exists (and if FM are still aiming for a 2010 release, it won't be long) then we'll hopefully see how good each GPU is in using D3D11. That's assuming FM don't plaster tessellation into every graphics test in 3DMark11.
Anyway, tessellation itself isn't necessary a weak area of AMD: note that I specifically said "geometry throughput", something which is independent of tessellation. The likes of the GTX 480 (and no doubt the 580) can setup 4 triangles per clock, whereas the best that AMD are currently offering is still only 1 triangle per clock. AMD could produce a GPU that's incredibly good at handling the process of tessellation (which itself isn't that hard) but until they do something about their setup engine, their performance in this area is always going to be capped the 1 tri/clk rate, especially when compared to NVIDIA.
Tessellation, in its current form in D3D11, can be done in D3D10; the improvement lie in the level of access to the control points in the patch (namely the hull shader and, to a lesser extent, the domain shader). Most of what they do was handled by relatively fixed function instructions in earlier D3D releases. Tessellation's a big thing at the moment because NVIDIA decided it would be.Most of the DX11 stuff can be done in DX10 except for tesselation (the relevant part)i believe and that is where AMD sadly gets crushed.
Last edited by Neeyik; 10-31-2010 at 03:36 PM.
Looks good if NVIDIA can deliver a 20% performance improvement in quantity on November 9th. Perhaps not coincidentally it will give NVIDIA something positive to talk about in their quarterly earnings report on November 11th.
I believe it was the Heaven benchmark that brought this whole thing to light.Arguably, tessellation has been a feature of D3D since v8 and it's only the level of control over the algorithms that's anything noticeable with D3D11; in other words, I disagree that it's a key feature compared to the introduction of the Compute Shader. Marketing/PR people would disagree with me simply because it's easier to promote tessellation ("look at all these bumps!!!1"") than the use of dynamic shader linkage ("those big shaders in games....well, they're now...umm, the same... but easier to handle!!!!1!"). We've seen no benchmark that really use SM5.0 yet, with all the new features and limits that it brings; until one such test exists (and if FM are still aiming for a 2010 release, it won't be long
Anyway, tessellation itself isn't necessary a weak area of AMD: note that I specifically said "geometry throughput", something which is independent of tessellation. The likes of the GTX 480 (and no doubt the 580) can setup 4 triangles per clock, whereas the best that AMD are currently offering is still only 1 triangle per clock. AMD could produce a GPU that's incredibly good at handling the process of tessellation (which itself isn't that hard) but until they do something about their setup engine, their performance in this area is always going to be capped the 1 tri/clk rate, especially when compared to NVIDIA.
Tessellation, in its current form in D3D11, can be done in D3D10; the improvement lie in the level of access to the control points in the patch (namely the hull shader and, to a lesser extent, the domain shader). Most of what they do was handled by relatively fixed function instructions in earlier D3D releases. Tessellation's a big thing at the moment because NVIDIA decided it would be.
The Heaven benchmark is a bit bonkers when it comes to using tessellation ("It's a flat surface! Give 1000 times more polygons!") but it's a test nonetheless. However, reviewers tend to use it inappropriately for when comparing geometry/tessellation performance: what one should do is remove as much of the pixel load as possible so that all of the remaining workload is in the form of geometry processing. Unfortunately that's not easy to do because you can't run it windowed and wireframe in the benchmark mode: all you can do is drop the detail settings right down and up the tessellation level as high as possible.
Another good test would be to look at NVIDIA cards in the benchmark where one keeps the core or shader clock at the default level but seriously reduces the rate of the other. Unfortunately I can't do that with my 460 (or more correctly, I can't see the option in Afterburner to unlink the clocks) but it would good if somebody could carry out these tests to see what clock reduction had the biggest affect on Heaven when heavily geometry loaded.
The GeForce GTX 580 mystery is clearing :-
http://www.nordichardware.com/news/7...-clearing.htmlGeForce GTX 580 will most likely be the most powerful single-GPU graphcis card on the market in 2010. Now that we have the complete specifications for NVIDIA's GF110 GPU it becomes clear that NVIDIA is attempting to do what it failed to with GTX 480 - going all in - to win the performance war against AMD.
While AMD is working on optimizing the Northern Islands architecture that beside better performance also brings new features, NVIDIA has an own agenda. To prepare its Fermi architecture for a new power struggle to maintain its position as the maker of the world's fastest snabbaste single-GPU graphics card. A big PR bonus NVIDIA has been able to use over the last six months.
If the increase in frequency and the so far unconfirmed information we've come across NVIDIA's work will result in a graphics card called GeForce GTX 580 that builds on a new optimized circuit called GF110. In this post we will try to summarize the information that we have so far and speculate in what this could mean for NVIDIA and the graphics market.
It's gonna be a monster as Cayman XT also appears to be, exciting times ahead and good to see the big players really locking horns, things have been rather dull of late.
Posting Permissions
Reply With Quote