|
Forgot to mention: details matter when analyzing timings, especially when comparing two very different systems like 8 and 9. For example, to make sense of timings you have to know the exact GPU card in use. Why? Here's an example: 8 will work with older generation, pre-Fermi GPU cards. 9 won't, as 9 is built around reasonably recent, Fermi or later GPU cards that can work with reasonably recent CUDA. If you try doing a mean curvature in 8 with an older GPU card, 8 could use it, but if it is pre-Fermi 9 won't use it (you can tell by looking in the Help dialog for 9) and instead will execute the job CPU-parallel. CPU parallelism in 9 is quick enough that it can often look like the GPU is being used. A radius of 5 for anything other than small data probably would be too much computation to mistake CPU parallelism for GPGPU, but if the data is smaller that could be an effect.
|