Get only the needed 2.3 dlls in beta downloads here, and get 190.38 driver from nVidia. Some small speedup at high angle ranges is observed with V12, will be experimenting to see if new compiler makes a better build, or not, for a while.
Are these the same dll's that are in the \CUDA\bin folder after you install the CUDA 2.3 Toolkit?
correct, the same old fftw library remains, and has nothing to do with cuda, but other parts of the application. Since the files have the same name, there are no app_info changes. Do make sure your driver is 190.38+ though.
N275GTX @666MHz, 896MB GDDR3 ram@2322MHz, 240 stream processors for $200N260GTX @655MHz, 1792MB DDR3 ram@2100MHz, 216 stream processors for $295So the questions are, are stream processors important to CUDA? Is less but faster memory better than more but slower memory? Is there something important to CUDA in the 275 vs 260?
I'm getting a video card, one of the less expensive types. It will be a MSI card, so my MSI mobo monitor will cover the GPU as well. I have two in mind - which will provide best CUDA performance?N275GTX @666MHz, 896MB GDDR3 ram@2322MHz, 240 stream processors for $200N260GTX @655MHz, 1792MB DDR3 ram@2100MHz, 216 stream processors for $295So the questions are, are stream processors important to CUDA? Is less but faster memory better than more but slower memory? Is there something important to CUDA in the 275 vs 260?
I'd definitely go with the 275. See this page for relative performance (GFLOPS column) of each at their default speeds: http://en.wikipedia.org/wiki/GeForce_200_Series . These GFLOPS numbers are not the same you see in BOINC, but they do scale well IME (8800GT, G94 9600 GSO, and G92 9600 GSO). I'm now running dual-GPU in linux with some of these and all works great!