Building a Better (Linux) GPU Benchmark

Some Test Cases

Our new benchmark wouldn't be very useful if we didn't have any test cases, would it? As we mentioned on the previous page, the average frames per second of a game might not reflect accurately what occurs in the game, even if the average results are very higher. Please note that we are anticipating a full benchmark of almost a dozen ATI and NVIDIA cards in the next few weeks - this is just a sample of what our benchmark utility is capable of. Both graphs below are taken with the same "radar" Wolfenstien radar timedemo.

The graph below is just an overlay of the above two graphs.


We can really see those dips in the graphs where some IO is occurring and bogging down the graphic cards. We have observed much unusual phenomenon on some of these demo graphs. One thing that we can immediately digest from the 5950 Ultra graph is the 90FPS ceiling - the card is more capable, but Wolfenstien is configured to limit the video card at 85FPS. Notice also that around the 73^rd second, the FX5600 hits a local minimum while the FX5950 hits the 85FPS ceiling. We will discuss these sorts of details more with the preliminary video card roundup.

Oh, yes, and of course, we have this graph to show you too.


Average FPS graphs are fine, but if we really want to convey some thorough and hard data sampling, we have to use average FPS in conjunction with line plots.