The problem with 3DMark, and other graphic benchmarks, is that none of the engines used are optimised. They simply chuck as many polygons and other effects at the card as is humanely possible, bringing it to its knees.
Real game engines DO NOT do this. Take the third test in 3DMark05. It looks amazing. It runs like crap. And that's because it's not an entire engine, it's just one small demo with unoptimised code designed for special effects. They don't design engines, they design demos. It's doing everything at full detail, all of the time, and real games just don't do that. They can't, if they need to be playable.
The devs at Futuremark don't have to worry about LOD, mipmapping, AI, net code or anything else - and all that means it's not in the least indicative of how a real game would render things.
In short, it's a false test. A couple of the tests in 3DMark03 still run waaaay worse for me than real, actual better-looking games (say, Doom 3, or Far Cry, or the EQ2 beta) - games that have actual interactive gameplay, instead of just a demo loop. Given that, why would I (or anyone else) rely on 3DMark?