Test Setup and Methods

We used the same CPU and mainboard as used to evaluate NVIDIA GeForce 8800 GTX for architecture investigations, but the rest of the component have changed. New games have been used, the big Dell 3007 has seen a bucketload of pixels over the last week or two and we've switched to using Vista x64 throughout. IHV's with subpar drivers for x64, take note: Vista x64 will be our performance testing OS for the forseeable future, at least when testing on Windows. The details of the test system follow. For all R600 testing we used the latest 8.37 press driver supplied by AMD recently behind the scenes.

Testing Methods

We use an updated set of our in-house tools, improving those we used to evaluate G80. Again, the architecture analysis revolves around instruction rate testing using our own issuers, to calculate throughputs and investigate execution limits of the shader core, whatever they may be, along with a mix of in-house and freely available tools to check performance elsewhere on the chip. The freely available tools come from George Kolling, Ralf Kornmann, Victor Moya, Mike Houston and pals at Stanford, and also the guys at iXBT/Digit Life, among others.

The rate issue tests are now a pair of binaries, running the same (pretty basic) HLSL shaders via D3D9 and D3D10 (where it makes sense and the hardware supports it), to check commonality of performance between the two D3D systems when looking for peak performance or bottlenecks or architecture intricacies. Whenever we're testing a theoretical peak rate and the application is windowed, we make sure the 2D clock profile matches the 3D clock profile so that we test the usual condition of fullscreen rendering. A new in-house filtering tester takes a look at multiple format filtering across 1, 2, 3 and 4 component surfaces.