How many samples per second would you like? Actually it can be made almost arbitrarily high: we simply make the samples black some fraction of the time (eg. 90%), and for the rest we bump up the brightness to compensate (eg. 10x) - this is the basis of Russian roulette, an essential method in unbiased rendering. Statistically this will converge to the same image, but the samples per second is magically 10x higher!neo0. wrote:Well, whatever it does I hope it is atleast as fast as other GPU renderers. In octane, I remember pulling 3 million samples per second in a scene that had over 10 million polygons..
Of course, this will take much longer to get a reasonable image, so you are actually shooting yourself in the foot if you care about producing a decent image in a given amount of time. The take-away point is that samples per second isn't a good measure of a renderer's "speed"; to give an analogy, if I declare "I can eat a thousand items of food per minute!" as measured with rice, and you're trying to match it by eating eggs, you would rightly object that the comparison is unfair.
I've made this point before in another thread:
My post here isn't directed at any particular renderers; for any two renderers their samples per second figures aren't directly comparable. With fewer samples per second you should expect the image to converge more rapidly - it's relatively easy to make a renderer that gives many samples per second under simplifying conditions, but how about one that delivers a clean image in reasonable time, using true unbiased rendering (a mathematical concept which has become a marketing term for something different), and in difficult scenes? There's an American saying for this: "all the speed in the world doesn't help you if you're headed the wrong way".lycium wrote:hmm there's no sense in attacking marketing stuff of octane or arion i think; for one thing it could be completely correct, eg. if you compare samples per second.
this of course doesn't take into account the kind of samples taken... for example, if you compare the results of:
1. typical gpu path tracing, deterministic light connection at end of path.
2. typical cpu path tracing, deterministic light connection at every bounce.
3. bidirectional path tracing, re-weighted sum of all scattering events.
i think many samples of one type are needed to get an equivalent result computed with the next; bidir makes theoretically optimal usage of a light path's information.
Indigo is aimed at delivering the highest possible image quality, and doing that as fast as possible. If you've done a comparison and find its performance to be lacking, we'd like to hear about it! We're always trying to improve our product :)