[Bf-cycles] Tiles sizes - impact on RAM and speed

Constantin Rahn crahn at vrchannel.de
Sun Feb 10 19:00:42 CET 2013

The magic number for fastest GPU rendering seems to be >= 128x128

I have posted some values in the benchmark thread on BA:
On a GTX580 128x128 is two times faster than with 64x64.

Am 10.02.2013 18:45, schrieb Brecht Van Lommel:
> On Sun, Feb 10, 2013 at 5:52 PM, Ole Jakob Skjelten <olesk at pvv.org> wrote:
>> Thanks a lot for that Brecht! My understanding is that *reducing* tile sizes
>> only has a positive effect with regards to reducing the size of the output
>> buffer, which is in most cases is a rather small part of the total RAM eaten
>> by Cycles, and thus, unless you you're in a desperate
>> I-only-need-10MB-more-RAM-to-render-on-my-GPU scenario, increasing the tile
>> size for speed benefits is the more likely scenario. Would you agree?
> Yes, I agree. Maybe if you have a ton of render passes, but otherwise
> you can quite safely set it quite large.
>> It's also my understanding that, using CUDA, keeping your tiles to multiples
>> of 16 pixels helps speed (or rather prevent wasting GPU cycles)? And
>> finally, does it help to operate with tiles that are multiples of the full
>> render resolution, or should one use same values for x and y direction
>> (square tiles) instead?
> Multiples of 16 might indeed help, I'm not quite sure if this true in
> general though and if this perhaps depends on the particular GPU.
> Multiples of the render resolution will help if you use big tiles,
> since otherwise you can get some smaller tiles to fill up the full
> render.
> Brecht.
> _______________________________________________
> Bf-cycles mailing list
> Bf-cycles at blender.org
> http://lists.blender.org/mailman/listinfo/bf-cycles

More information about the Bf-cycles mailing list