Hmmmmm. Maybe I take that back. Reading from "CUDA Programming Guide Version 3.0" (http://developer.download.nvidia.com/compute/cuda/3_0/toolkit/docs/NVIDIA_CUDA_ProgrammingGuide.pdf), I see:

For devices of compute capability 1.x, a multiprocessor consists of:
8 CUDA cores for integer and single-precision floating-point arithmetic
operations,
(Appendix G.3, page 142)

For devices of compute capability 2.0, a multiprocessor consists of:
32 CUDA cores for integer and floating-point arithmetic operations,
(Appendix G.4, page 146)

So maybe David's way is the only officially-sanctioned one. But if you have to derive it from the Compute Capability, shouldn't the test be (prop.major>=2)?

The 'compute capability' for a Fermi is 2.0: it's the 'CUDA version' which has gone up to 3000 or above. That's a software number (from '__cuDriverGetVersion'), not hardware (from '__cuDeviceComputeCapability').

Reporting from GPUGrid:

13-4-2010 9:58:53 Starting BOINC client version 6.10.45 for windows_intelx86
13-4-2010 9:58:55 NVIDIA GPU 0: GeForce GTX 480 (driver version 19741, CUDA version 3000, compute capability 2.0, 1536MB, 1345 GFLOPS peak)

my 470 is showing as 726 GFLOPS up from 128 (going from .43 to .45 clients).

certainly not as cool as my 2TFLOPS from each 5850.....

that look right?