deflix wrote:Are you saying that to achieve an image at a given SL it is just as efficient to render in Coop as it would be to render seperate regions on seperate nodes and put them together?
<technical rambling>
This is actually central to the "unbiased" part of "unbiased renderer".
Unbiased renderers produce pixel samples which are averaged together to produce images with less and less noise. As long as two separate renderings were done with different random seeds, averaging them together is no different than doing a single render for twice the time.
That is the theory.
Now, granted, in practice, it may actually be the case that you see a small speed boost if you render, say, four tiles of 1000x1000 instead of four 2000x2000 images. The reason for this is not theoretical, but practical, and that is because if the tiles become small enough, they may lead to better cache coherency and reduce the degree to which Maxwell is waiting on memory accesses.
As an extreme example, suppose you rendered an image so large that it was actually larger than the core memory and had to use your systems virtual memory to store it. For example, suppose a machine had only 256 megs of memory in it, and you were rendering a 256 megabyte image. Obviously it cannot fit that, plus the OS, plus the scene, plus the textures, etc., so it is going to be swapping out to disk all the time.
This would
drastically reduce the rendering efficiency, because now there's actually hard drive reads and writes involved in updating the image.
As a result, rendering that image as four 64 megabyte images would dramatically improve the rendering speed, because now there would be no VM swapping.
Make sense?
The same rules apply to CPU caches and main memory, just to a much less dramatic extent. So whether or not it would be an important speed win to do small tiles is something that we'd have to do extensive testing to determine. My intuition is that the tiles would have to be small, like 512k or one megabyte or something, before you would see a worthwhile speed improvement.
Since Maxwell is presumably keeping things around in floating point color, we're talking chunks of 128x128 or 256x256 at most.
In general, regardless of all other concerns, the entire improvement possible from cache optimization of this form would depend on the amount of time Maxwell actually spends waiting for memory accesses to read, modify, and write back pixel color computations. It is entirely possible that this amounts to a negligible amount of time (2%, 3%, 5%, etc.), in which case no matter how much of a win the tiles were, you're not going to see a speedup that you actually care about.
Since I don't work for Next Limit, I don't know what that percentage is, but someone there probably does
</technical rambling>
- Casey
i am just thinking about upgrading from 4 to 8 gb on the nodes..
- By Mark Bell
- By Edward Leibnitz
- By Forum Moderator