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

[Brecht Van Lommel]

Hi,

I don't have a simple answer for this either, all I can suggest is to
just try it out. Basically you have to balance a number of things and
it's difficult to make predictions because it depends on the GPU and
scene being rendered.

Bigger tiles gives more memory usage especially if you have many
render passes enabled, since it has to fit the entire tile in memory.
The other disadvantage is that if you have multiple GPU's, you might
end up with one GPU working on a tile while the other is doing
nothing.

With smaller tiles the disadvantage is that the GPU might not be fully
utilized, each pixel is handled by one core, and GPU's have hundreds
of cores. Further GPU's tend to work more efficiently when they have
multiple threads per core. The other disadvantage with smaller tiles
is that this means more GPU/CPU communication which can also lead to
the GPU being underutilized.

I think generally the advice is to use tiles sizes of 256x256 and up
for GPU's, but I don't have a single formula that I can recommend.

Brecht.

On Sat, Feb 9, 2013 at 7:13 PM, Ole Jakob Skjelten <olesk at pvv.org> wrote:
> Hi!
>
> I've spent quite a bit of time recently setting up a GPU hardware FAQ for
> Cycles, based mostly on the recurring questions from the support forum at
> BlenderArtist. I feel I got most of it covered, but one type of questions
> keeps coming up that I do not have a good answer for, and that is the impact
> Cycles tile sizes has on total VRAM and rendering speed. I've seen a lot of
> empirical testing, but it's hard to transfer this into some solid general
> advice.
>
> I suppose I'm trying to answer the questions "which tile size gives optimal
> speed given my scene and hardware?" and "how does tile size impact VRAM
> requirements?" (which also answers the questions "I get an out of memory
> error, can I do something about that?").
>
> I was hoping there was some sort of formula I could use to understand the
> relationship between tile size and VRAM usage, and the impact on number of
> calculations required for a full scene. I've asked around everywhere I could
> think of, and though people have a lot of anecdotal evidence one way or
> another, I've yet to get any closer to the underlying technical details. But
> perhaps there is someone on this list that could settle this once and for
> all? (I promise to distribute the answer broad and wide ;)