GPU Shootout with Unreal Tournament 2003 - July 2002

The Benchmarks

We first introduced the Unreal Tournament 2003 benchmark in our recent Parhelia review, the build we're showing off today doesn't differ too much from what we demonstrated in that review.

We tested at five resolutions, each under two different detail settings. We chose the highest detail setting offered by the game, with everything set to the maximum level to provide the best possible image quality. We also chose a medium detail setting, which turned off detailed textures, turned on 16-bit color, lowered the texture detail and turned off deco layers. The reason for the medium detail setting was to compare those cards that weren't playable at the higher settings.

We ran all of those benchmarks under two levels: DM-Antalus and DM-Asbestos. We used the Antalus deathmatch level in our Parhelia review, but the Asbestos level is a newcomer to our reviews. The main difference between the two levels is that Antalus is a mostly outdoor map while Asbestos is entirely indoors and thus provides much higher frame rates.

We also had Epic whip up a test that would specifically isolate and test memory bandwidth over anything else. The reason we wanted a purely memory bandwidth test was to put numbers to John Carmack's latest words about Matrox's Parhelia:

"The performance was really disappointing for the first 256 bit DDR card. I tried to set up a "poster child" case that would stress the memory subsystem above and beyond any driver or triangle level inefficiencies, but I was unable to get it to ever approach the performance of a GF4. " - John Carmack, 6/25/2002 .plan update

The resulting test we came up with was a large room that was tiled with three 2Kx2K textures, uncompressed, across the height and length of all of the walls. The end result wasn't a pretty level at all, but it was perfect for stressing the memory subsystem of the cards. We couldn't run this test on all of the cards, as those with only two rendering pipelines (and two texture units per pipeline) or less than 64MB of memory wouldn't either complete the test or would swap to main memory entirely too much.

In order to make this comparison a manageable task we limited the tests to a single system, but if you want us to we will spend time this week running our CPU scaling tests on a number of the cards. As usual, your wish is our command so just let us know.