Games Benchmarks - DirectX
Max Payne
For the first games benchmark we'll use the DirectX title Max Payne.
| FPS | 640x480 | 800x600 | 1024x768 | 1280x1024 | 1600x1200 |
| GeForce 4 Ti4600 | 159.0 | 155.7 | 121.7 | 79.0 | 57.1 |
| GeForce FX | 150.6 | 150.9 | 150.8 | 136.5 | 104.2 |
| GeForce FX (400/400) | 151.6 | 151.2 | 147.4 | 115.5 | 85.6 |
| % Diff from Ti4600 | 640x480 | 800x600 | 1024x768 | 1280x1024 | 1600x1200 |
| GeForce FX | -5% | -3% | 24% | 73% | 82% |
| GeForce FX (400/400) | -5% | -3% | 21% | 46% | 50% |
At low resolutions all the boards start out limited by the CPU or system, and their on board Geometry processing is not able to overcome other bottlenecks on the system, so all three setups have roughly the same performance. Oddly, though, at this low resolution GeForce4 has a small edge in the performance; this may be due to a slightly older and more optimised code set for GeForce4 in the drivers, or it could be because GeForce FX 5800 isn't geared toward low resolution running.
As we scale up through the resolutions we begin to see some differentiation between the boards and speeds by 1024x768, with the GeForce4 already showing fill-rate limitations. The GeForce FX at 400/400 and 500/500 speeds are beginning to run out of fill-rate at 1280x1024, but this is obviously more so with the FX 5800 running at 400/400. By 1600x1200 the 500/500 FX has an 82% performance advantage over the GeForce4 Ti at its default IQ settings.
| GeForce FX 5800 Ultra | 640x480 | 800x600 | 1024x768 | 1280x1024 | 1600x1200 |
| Normal | 150.6 | 150.9 | 150.8 | 136.5 | 104.2 |
| 8x Aniso | 145.1 | 145.0 | 140.3 | 112.6 | 85.6 |
| 4X FSAA | 149.4 | 147.6 | 121.1 | 80.5 | 44.3 |
| 8x Aniso + 4X FSAA | 142.5 | 136.1 | 103.0 | 68.1 | 38.9 |
| % Difference | 640x480 | 800x600 | 1024x768 | 1280x1024 | 1600x1200 |
| 8x Aniso | -4% | -4% | -7% | -18% | -18% |
| 4X FSAA | -1% | -2% | -20% | -41% | -57% |
| 8x Aniso + 4X FSAA | -5% | -10% | -32% | -50% | -63% |
Utilising 8X Anisotropic Filtering (AF) under the default driver mode of 'balanced' in the 5800 Ultra incurs a performance penalty of, at worst 18%, due to the increased overhead of the extra texture sample fetches. In this instance the board is still performing over 100FPS at 1280x1024 and well above 60 FPS at 1600x1200.
Using 4X FSAA with the 5800 Ultra accounts for a worst case performance hit of 57% in this title. Although GeForce FX features colour compression, it is not entirely able to offset the increased bandwidth demands of MSAA. Looking at the fill-rate graph we can also see a drop in fill-rate performance at 1600x1200 which would indicate that onboard memory space is at a premium here, and some textures are being addressed over the AGP bus. At 1280x1024 the performance is still well about 60 FPS on average, dropping a much lower than 60 FPS at 1600x1200.
With both 4X FSAA and 8x AF enabled the performance penalty rises a little, but as is often the case with multisampling and Anisotropic filtering the performance penalty is not cumulative as each of the features are stressing different elements of the pipeline. With both these features enabled under this benchmark the 5800 Ultra is still a little over 60 FPS at 1280x1024 and over 30 FPS at1600x1200, with a worst case performance penalty of 63%.
