For CUDA 2.3 (or 2.2, if it’s not too late), what about adding some functionality into the Windows Performance Counters:
You could also probably auto-install a template into the Vista/Server 2008 resource monitor (Task Manager -> Performance Tab -> Resource Monitor). It’s got a bunch of neat stuff in there to monitor which programs are using CPU/Disk/Network/etc., and you can add custom panes to it as well. It would be nice to have a GPU monitor that registered the amount of GPU usage, I/O bandwidth usage, memory access, and so forth.
If you didn’t add the template, at least you could always access it through the same API, and I’m sure the people that are running some HPC clusters with Tesla cards might appreciate being able to monitor the performance of the cards remotely.
Even beyond the development sector, I could see gamers (and perhaps even the average joe using GPU-accelerated apps) wanting to have access to this information. Perhaps it could be included in an upcoming driver release (since it would be part of the driver, and not CUDA specifically)?