The original F.E.A.R had a great benchmark in my opinion. It was relatively quick, gave a good sampling of different things to strain the card that would be realistically encountered in the game, gave some nice quick stats at the end (min, max and average framerates), and most importantly of all, each benchmark run was identical to the last in terms of what occurred.
Partial tangent, I hate games where you choose an "ultra" or similar preset (basically whatever is their highest offered global preset) and there are a few settings that can still be pushed up one or two more levels. If I pick the preset for the highest thing, I expect every setting to be at it's highest. I usually just manually set everything myself because of this.