Sign in with
Sign up | Sign in

Vertex/Pixel Shading Performance

Nvidia GeForce GTX 260/280 Review
By , Florian Charpentier

Since they should be able to draw some improvement from the near-doubling of the number of ALUs, first let’s see how the cards perform with the Rightmark Vertex Shaders test (all with the same performance).

gtx 260 280

Surprisingly, and despite different adjustments and successive tests, the GTX 280 not only posted poorer performance than the 9800 GTX, but fell 12% short! Nvidia also got the same results, and was able to change that only by enabling 4X antialiasing – surprising for a geometry test. We should note, however, that while processing power (and consequently transformation power) has increased strongly, the setup engine hasn’t been changed. As with the 9800 GTX, it’s capable of generating only one triangle per cycle. The latter’s frequency advantage (675 MHz compared to 600 MHz), might explain the difference.

gtx 260 280

As usual with RightMark 2.0, the first shader showed no improvement with the new card, whereas the second showed a 25% gain.

Though we’ve already devoted a lot of space to advanced pixel shaders via the preceding tests (arithmetic in particular), let’s look at simpler shaders, with the Fillrate Tester per-pixel lighting test we’ve been using for four years now.

gtx 260 280

We’ve come a long way since then! However, we really could have expected more than only a 40% gain from the GTX 280. Similarly, and without showing all the extensive results for ShaderMark (which uses Pixel Shader 3.0), the increases remain at between 20 and 26% for the last six shaders, and it did no better than 43%.

All the results are surprising, and show the gap between the theoretical increase in power (which should show up with Vertex and Pixel shaders, even old ones, in particular) and the actual gains in applications. We’d bet that the drivers aren’t fully optimized yet, and don’t forget that even when using a very specific benchmark, it’s still difficult to isolate a specific aspect without being influenced by the rest of the pipeline, especially with current architectures.

React To This Article