You'd hope that it'd scale well in a benchmark, tbf. Most instances of poor scaling in the real world are to do with bottlenecks with the CPU or VRAM bandwidth. Most instances of zero scaling are driver related. I've totally seen the light with regards to the flaws in throughput-based comparison, though. As such, I think 3Dmark, which only tests throughput (as opposed to latency), is kinda useless as anything other than e-peen.