When I only had 2 GPU WUs I noticed that one was on the slow GPU rather than both being on the fast GPUs. The Event log reports:

20/02/2012 21:40:50 |  | NVIDIA GPU 0: GeForce GTX 460 (driver version 28562, CUDA version 4010, compute capability 2.1, 1024MB, 684 GFLOPS peak)
20/02/2012 21:40:50 |  | NVIDIA GPU 1: GeForce GT 430 (driver version 28562, CUDA version 4010, compute capability 2.1, 512MB, 179 GFLOPS peak)
20/02/2012 21:40:50 |  | NVIDIA GPU 2: GeForce GTX 460 (driver version 28562, CUDA version 4010, compute capability 2.1, 1024MB, 684 GFLOPS peak)
20/02/2012 21:40:50 |  | NVIDIA GPU 3: GeForce GT 520 (driver version 28562, CUDA version 4010, compute capability 2.1, 512MB, 104 GFLOPS peak)

Even when I exited and restarted BOINC the first two in the list were being used rather than the 2 GTX 460s. The above order is the order in which cuDeviceGet() returns them so my guess is that BOINC should be changed to sort based on the GFLOPS peak and use the fastest ones before the slower ones.
I know the two slow ones don't do much but my having them it effectively means I get 800 WUs for each GTX 460 to munch rather than 400 so can survive longer when the servers are away.