Is it possible on either the geforce or quadro series?
Yes, but why this question is in CUDA forum? =)
Possibly, an underclocked high end gpu could provide a better performance/watt ratio than a mid-level card. No one outside of GPGPU cares about GPU performance/watt.
Reviews I’ve read of NVIDIA cards indicate that the G90+ devices are significantly better performance/watt than, say, the 8800GTX. The 8800 GTS 512MB, in particular, is quite efficient, if i recall correctly. The GTS is significantly cheaper than the GTX,as well.
That said–I would doubt that most GPGPU people care about performance/watt as much as other considerations when buying a GPU. I myself, for example, run memory-limited algorithms that led to me buying Tesla Computing Solutions instead of 8800’s. I would bet that other people place greater emphasis on clock speed (thus buying 8800 Ultras) and others might be bandwidth limited (8800GTS 512).
For mobile computing, performance/watt rather than price/performance can drive purchasing decisions.
You’d be better off taking a look at some of our lower-clocked GPUs than trying to slow something fast down. Check out the mobile GPUs or lower powered ones. There are over 30 models of CUDA-capable GPUs that vary by number of cores, memorory size and clock frequencies, with excellent perf/watt. A single cuda binary can run on all of them, though you may want to account for available memory.
More than power consumption, I am really interested in long term reliability and thermal management. Accelerated life studies would be great to have in the public domain. For example, if I started a while loop that calls a computationally intensive matrix multiply kernel and came back to the same laptop a year later, how likely is it that some of the transistors will be effectively “fried”?