How valid is Tessmark?

[Red Hawk]

Prior to 2013, Anandtech used the DirectX 11 Detail Sample benchmark for measuring tessellation performance:

http://anandtech.com/bench/GPU12/423

And also used the more "real-world" test of Unigine Heaven:

http://anandtech.com/bench/GPU12/409

However, it got to the point where the detail sample wasn't really pushing top-end DX11 cards any more, either from Nvidia or AMD. So in 2013 Anandtech started using Tessmark:

http://anandtech.com/bench/GPU13/596

Now the 2013 results made me raise an eyebrow. How was it that a 7970 GHz Edition lagged so far behind a humble Geforce GTX 660? But I guess it made some sort of sense; the 7970's theoretical tessellation isn't technically greater than a 7870's, as they have the same geometry engine. Maybe Tessmark simply makes the software techniques that AMD used to improve performance useless. But when you look at the 2014 results, things get really confusing:

http://anandtech.com/bench/GPU14/841

The 290X is barely better than the 7970, and still below the GTX 660. The 290X has twice the geometry engines and thus tessellators as the 7970 does. Theoretically -- and the Tessmark benchmark is essentially a theoretical, not real-world, test -- it should have twice the tessellation power of 7970. But the test doesn't reflect that. On the other hand, games which make relatively thorough use of tessellation are fine on the 290X:

http://anandtech.com/bench/GPU14/852

In Metro: Last Light, the 290X only comes up short of the GTX 780 Ti, and the GTX 660 comes in just behind the 7870.

It seems to me that Tessmark does not meaningfully measure real-world tessellation performance for whatever reason and ends up misrepresenting the capabilities of AMD cards.

 

[lyssword]

It sounds plausible, because AMD cards are known to be weak at 64x tess performance. I don't think many games use 64x

 

[Hitman928]

http://www.computerbase.de/artikel/grafikkarten/2013/amd-radeon-r9-290x-im-test/13/

 

[zlatan]

Tessmark is valid. The problem is that AMD only use one setup pipeline in their openGL drivers, no mater that some GPUs have two or four pipelines. In D3D all pipelines work.

The DirectX SubD11 sample is the best test to do synthetic analysis for raw tessellation performance. But these result won't represent the real world scenarios, because in a game there will be many other limitations such as overshade and bad rasterizer efficiency.

 

[Arkadrel]

its not reflective of real world gameing, thus not very valid.

 

[Red Hawk]

Tessmark is valid. The problem is that AMD only use one setup pipeline in their openGL drivers, no mater that some GPUs have two or four pipelines. In D3D all pipelines work.

The DirectX SubD11 sample is the best test to do synthetic analysis for raw tessellation performance. But these result won't represent the real world scenarios, because in a game there will be many other limitations such as overshade and bad rasterizer efficiency.

Well...on the one hand, it's pretty bad of AMD that their drivers don't use more than one pipeline for OpenGL tessellation. If there were to be an OpenGL game (on, say, Linux with SteamOS) AMD really would lag behind Nvidia that badly. That said, the vast, vast majority of games which have advanced enough graphics engines to use tessellation are DirectX 11 games. As long as that's the case, Tessmark will be no good as a practical benchmark for comparing tessellation between Nvidia and AMD cards.

 

[Silverforce11]

Synthetics are just for amusement. Always have been. Good for e-peen suicide OC runs! Even the widely used 3dMark is very unrepresentative.