Originally Posted by Ph0en1x
What is the calculation speed difference between calculation using only CPU vs calculation with CUDA graphic card ?
Just to be clear... currently, the publicly available version of SnG Sovler does *not* support CUDA yet (even though you might have noticed that a CUDA DLL is part of the installation). But that said, I can tell you a few things based on some development builds...
On my dev machine with an 2.2Ghz 8-core Xeon CPU and a GTS450 GPU (192 CUDA cores, 1.5Ghz), I see a speed increase of 5-8x when using the GPU.
This seems to be in line with the theoretical performance of each device based on their relative GFLOPS potential (72 vs 576, respectively), but I still believe there is more to be had from further optimizing my CUDA implementation. A real-world GPU/CUDA application has a much better chance of actually approaching maximum theoretical performance than its CPU-based counterpart. So really, I think GPU SnG Solver should ultimately perform better than it would look "on paper" relative to a CPU-only version.
In other words... SnG Solver performance is basically proportional to how many FLOPS you can give it.
For typical modern CPUs: FLOPS = 4 * # physical
CPU cores * CPU frequency
For "Fermi" GPUs: FLOPS = 2 * # CUDA cores * GPU frequency
For "Kepler" GPUs: FLOPS = # CUDA cores * GPU frequency