Quote - I have now tested both the GTX580 and the GTX680 with Octane. I can confirm that the GTX580 is faster (consistently, about 15% with Octane 1.0 RC3), despite the GTX680 having 3 times the number of CUDA cores. Within a series (5xx or 6xx), speed should be proportional to the number of CUDA cores.

Mmmmm, that's strange.  Are you testing on one PC or two?  If one, are both cards in the PC at the same time, or are you physically swapping them?  Are you using identical camera settings for the test scene?

Actually, that is what would be expected. There are some significant changes between the 500 (Fermi) and 600 (Kepler) series cards that can have a big effect on Cuda performance.

For one, the Fermi cards process 2 instructions per clock cycle, the Kepler cards process only one instruction per cycle. This means if all other things were equal, at the same clock speed, you would need approximately twice as many Kepler cores to match the per clock cycle performance of a Fermi card. The second big difference between Fermi and Kepler is the ability to efficiently perform floating point calculations. The reduction in transistors per core in Kepler came with a reduction in the efficieny for processing floating point calculations. Any process that relies heavily on floating point calulations will run slower.

For gamers the new Kepler architecture is much faster, as it is optimized for shading rates. The plus for people using Octane, or another GPU render engine, is that the new achitecture can handle a greater number of texture maps.

__________________________________________________________

My Rendo Gallery ........ My DAZ3D Gallery ........... My DA Gallery ......