Hi,
I’m now using titan X and CUDA 5.5 for cuda programming.
However, I’ve faced a problem which is low performance.
when I run sample ‘nbody’ via CUDA Sample Browser ‘run’ botton, it seems to be good performance.
But, when I executed nbody code via visual 2010 manually, the performance was really bad.
I think that Sample Browser’s Performance is not for my GPU…
More information the performance is below.
via CUDA Sample Brower
228 fps, 133 BIPS, 2800 GFLOP/s
via visual 2010 manually
1.9 fps, 1.2 BIPS, 23.5 GFLOP/s
**
Also, CUDA7.0 has only manual running, so I met the same problem with CUDA5.5.
really low performance…
Does anybody have an idea?!!
How can I fix it to get high performance?
Make sure you are compiling for a release build, not a debug build. In particular, the nvcc command line should not include the switch -G.
I have not run the nbody sample app in a very long time, but as I recall the app has a command line switch that you need to specify to turn on benchmarking mode (which is disabled by default), otherwise performance suffers noticeably due to CUDA /OpenGL interop overhead.