Proof of what I said in my previous blog, from an unexpected place

Beyond3D has published an analysis of the Fermi architecture. The interesting part there is that they did a thorough analysis of the triangle setup/tessellation performance, and compared it with the Radeon 5870.

In a nutshell they concluded two things:

  1. The triangle setup of the Radeon is more efficient than the GeForce.  In fact, the 400-series actually has barely faster triangle setup than the previous generation (they claim the GeForce is actually artifically limited, and the Quadro series have better triangle setup performance, making full use of the parallel ‘PolyMorph’ setup). The Radeon is considerably faster than either one.
  2. The tessellator of the Radeon is serial, and has very limited throughput, causing a huge bottleneck, where the parallel ‘PolyMorph’ tessellation setup of the GeForce scales much better.

This supports what I said in my previous blog: The Radeon’s problem is not that small triangles aren’t efficient. In fact, the Radeon is more efficient than the GeForce at all triangle sizes, by a margin. So technically the Radeon is actually better at rendering small triangles. So Richard Huddy’s claim of triangles smaller than 16 pixels hurting performance on Radeons is nonsense. It would hurt GeForces more than Radeons, and certainly doesn’t explain the performance differences we see in tessellation benchmarks.

The problem is in the limited throughput of their tessellator, as I already said. And this is pretty much independent of triangle size.

So there you go, Richard Huddy. Even Beyond3D agrees with me, and debunked your ‘explanation’ of why Radeons perform poorly in the tessellation benchmarks. Not because nVidia is being unfair, and certainly not because of the triangle size used. No, it is the poorly performing and scaling serial tessellator setup in your architecture, as I’ve been saying all along (yea, too bad for all those people who tried to claim that I didn’t know what I’m talking about, and how I should try discussing this at Beyond3D… Consider yourself owned).

From here it looks like Richard Huddy doesn’t know what he’s talking about. I suggest he goes looking for a different job.

This entry was posted in Direct3D, Hardware news, Software development, Software news, Uncategorized. Bookmark the permalink.

2 Responses to Proof of what I said in my previous blog, from an unexpected place

  1. Pingback: Remember caveman-jim? | Scali's blog

  2. Pingback: AMD tries to do more damage-control for tessellation | Scali's OpenBlog™

Leave a comment