If you've done any benchmarking yourself you know this is complete bullshit. Performance improvements/regressions had in one title on an updated driver/kernel needn't translate into performance improvements/regressions on another title. This is a FACT.
Originally Posted by ua=42
Another fact is that people don't care too much about precise reproducibility. They care about some rough idea of how a game THEY ACTUALLY PLAY performs on hardware roughly comparable to what they have. Add a disclaimer to the benchmarks saying that they were not automated, though most probably wouldn't even care! But just benchmark games people care about!