The 3DMark03 plot thickens, again


— 3:59 AM on November 13, 2003

Beyond3D has done a comprehensive analysis of 3DMark03 build 340 performance that goes so far as to suggest that NVIDIA has broken all of its driver optimization guidelines when it comes to 3DMark03:

One the face of it here, the optimisations applied here appear to break all three guidelines - the correct images do not appear to have been produced with the 52.16 drivers, pre-computed state in the form of shader replacements appear to have been used (if this were part of a generic optimisation then it should not have been defeated by simply altering the order of instructions), and hence these optimisations would be specific to this benchmark only. Given that we first heard talk of NVIDIA's optimisations policy during the Detonator 4x.xx release of drivers there would appear to have been ample time to have removed these types of optimisations by now.
Yes, this story refuses to die. But it gets better. HardOCP has a response from NVIDIA on the Build 340 controversy in this benchmarking editorial. NVIDIA claims that the latest 3DMark03 patch is specifically designed to defeat the Detonator driver's unified compiler, which appears to be responsible for not only instruction re-ordering, but also shader replacement.

Thus far, I've been pretty impressed with NVIDIA's run-time compiler; intelligent instruction re-ordering and shader replacement are fine by me as long as rendering output doesn't change. However, I'm not so keen on FutureMark specifically targeting NVIDIA's compiler to create an apples-to-apples environment where all code optimizations are forbidden.

Tip: You can use the A/Z keys to walk threads.
View options

This discussion is now closed.