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

[Martijn Berger]

I did some benchmarking with 3 versions of the cuda toolkit on my box.

My interpretation of the results is that it could be good to move to cuda
5.0.
But more importantly 24 max registers seems to be sub optimal in the scenes
I have tested with.

It could maybe be good to see if the default for sm_21 and sm_20 cards
could be moved to 32 as it is for sm_3x currently.

Also note that I tested with the release 2.67b linux from b.org.

Martijn
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.blender.org/pipermail/bf-cycles/attachments/20130612/5bd86490/attachment.htm 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: CUDA_results.pdf
Type: application/pdf
Size: 108842 bytes
Desc: not available
Url : http://lists.blender.org/pipermail/bf-cycles/attachments/20130612/5bd86490/attachment.pdf