[Bf-cycles] 3 way cuda performance comparison on sm_21 and sm_20 cards

[Jürgen Herrmann]

Hi, 

 

I did extensive tests on windows with different cuda toolkits and different
builds (VS 2008, 2012 and MinGW) lately to see if I can find a reason why
Cycles is so much slower on Windows/MSVC builds.

My results can be reviewed here : http://download.shadowrom.de/Benchmark.pdf

We could speed up CUDA performance on all OS if we use nvcc compiler
optimization switches and some optimization in the cuda kernel code.

 

I’ve also tried to profile cycles with “Very Sleepy” on Windows MSVC builds
too. I am still experimenting on how to speed up the time critical parts for
M$ compilers 


 

/Jürgen

 

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.blender.org/pipermail/bf-cycles/attachments/20130614/6e055f60/attachment.htm