Theoretical performance
Theoretical numbers are all fine and well on paper, but how do the cards pan out when tested? We'll be looking at the cards' performance via the theoretical testing section of 3DMark 2003. Whilst the benchmark itself has come in for some criticism, the theoretical section can give us some much-needed information with respect to possible performance.
Tests were run at 1024x768x32. The difficulty of the tests should ensure that the onus is put on the cards and not on the subsystem. Looking at single-texturing fillrate first.
Remember that both the NVIDIA cards drop down to outputting 4 pixels per clock cycle when outputting a standard pixel (with a colour value). The ATi cards, though, have a standard 8 rendering pipelines. That's why the 9800 Pro takes this benchmark with relative ease. The 5900's pretty efficient here.
Running multi-texturing, though, brings into play the NV cards' 2 texture units per pixel pipeline. They comfortably take the multi-texture benchmark.
Advanced pixel shading helps games develops to produce far more realistic-looking models. How does the FX5900 Ultra fare?
The pixel shader 2.0 ability seems to be have increased on the FX5900 Ultra, but even if you factor out the clock deficit it has in relation to the FX5800 Ultra. the shading ability is nowhere near double. We're also surprised to see it trail the Radeon 9800 in this test. Expect a nice, shiny driver release to boost performance across the board.
Vertex shading now.
Hmmm, not so good here. The multiple ogres take their toll on the FX5900's vertex array.
