There is an industry standard benchmark for measuring overall system performance that we (as well as everyone in the, uh, industry) use for our own benchmarking, competitive analysis, etc.
It's made up of multiple suites, and one of the suites is a 3D subsystem test, but it is 3D in name only. It loads a scene in a popular 3D application...and renders it to a file. Basically, it is a CPU stress test. Great for processor scaling...not so great for 3D graphics, as the benchmark result scales with processor frequency and cache and not with the graphics chipset.
I'm working on developing a new 3D suite, and one of the subtest will be a cityscape with airships, roads, etc. It's in its infancy, but here are some preliminary shots. No lighting, no textures, and these are only four city "tiles."

My plan is to duplicate these tiles with rotation and some tweaking, as well as add some larger buildings and skyscrapers to break up the skyline.
Here is a shot from within 3D Studio MAX. I'm using the built-in Polygon Counter and I've highlighted one of the original tiles. Each tile is 27080 polygons, with the set of four reaching 108320 polygons.

I'm not sure how close I'll get to this level of buildings, but I definitely want to make an effort on texturing so they'll hold up to observation while the scene is playing.