slow vray Next vs 3.6

[^Lele^]

I, Vlado, or support can only find out what is happening by looking at an actual scene.
Feel free to send it to me (to the mail in my signature, below) so i can have a look, and then suggest a course of action.

[Karonte]

Ok. thanks.
I didn't want to disturb you. I just looked for a quick and dirty trick...anyway I will send you .

[^Lele^]

Ok. thanks.
I didn't want to disturb you. I just looked for a quick and dirty trick...anyway I will send you .

It's very ok, no worries.
There could be one or a few dirty tricks.
But i'll know for sure once i'll have looked at the scene.

[JonasNöll]

The images posted above show an average noise level which is over double for the 3.x than it is for Next.
Both are ofc below the visible threshold, but that doesn't mean there are no differences.

In fact, once one takes into account noise levels, and normalises time, the emerging scenario is rather different.

In the linked contact sheet, you'll want to check the value of average noise level, in the middle row of values, the last of four (V).
0.001047 Vs. 0.002697 is a difference of 2.57X, while the rendertimes were slower by a factor of 2.04X, making V-Ray Next somewhere around 25% faster for this scene (hello marketing claims.)

Notice that doing these kind of assessments with integer images is bad, and it's made worse by using Photoshop.
You'll want Nuke, Fusion, or equivalent FP imaging tool, with FP imagery, to make sense of this stuff.

the contact sheet is here.

p.s.: for the scene you posted, instead, the noise is visually much lower for Next/v5. An A/B in the VFB shows it even without need for post measurements.

Hi, thanks a lot for your clarification, very interesting. So if I understand correctly that means that Next by default might need longer to render because it calculates the noise treshold not exactly the same like in 3.6 and this way results in a slower rendering but higher quality result? I agree that in my scene with the teapot the background seems more noisy in 3.6 so that would support this.
However now I tried another approach, please correct me if there are logical issues:
I disabled the max subdivs in the Bucket image sampler and render everything with 1 Min subdivs, this way we wouldn't need to guess about different noise tresholds and so on, right? Then because the rendering goes pretty fast I increase the resolution to for example 5k and render with 3.6 and Next and also here Next renders consitently slower. 59s (3.6) vs 01m:14s (Next) in my case.
This would indicate that the difference does not only come from the different noise tresholds if I understand correctly or? As what Vlado mentioned that Next calculates the scene more accurate and this way increases the rendertime?

As for the noisetreshold it would be interesting to know which settings would roughly correlate to the same settings in Next. For example I rendered 1 Min 8 Max NoiseTreshold 0.01 in 3.6 then how would I need to set up this in Next to reach the same result? Or is it not so eaily comparable?

[kosso_olli]

We found out that a noise threshold of 0.005 in 3.6 roughly correlates to about 0.008-0.01 in Next.

For example I rendered 1 Min 8 Max

These settings are not good. You are choking the adaptive algorithm by such a low Max value.

[JonasNöll]

We found out that a noise threshold of 0.005 in 3.6 roughly correlates to about 0.008-0.01 in Next.

Alright, that's interesting to know, thanks! I will give it a try

[JonasNöll]

Ok so I tried with different noise tresholds and tried to find an equilibrium between rendertimes. In this simple scene I found that noise treshold of 0.008 in 3.6 roughly matches the rendertime than in Next with 0.01.
However I must admit that even with lower treshold of 0.005 for example the grey background looks still noisier than in Next with 0.01. While other parts of the teapot look smoother in the 3.6.
Maybe because so many things changed between 3.x and Next this approach anyway possibly doesn't make sense. Anyway I though to share this here in case someone finds it useful. I don't have any professional tools to analyse the amount of noise in each picture and can just evaluate it visually. Would be interesting to know though at what treshold the average noise in 3.6 reaches the average noise in Next. In case somebody want to do that HERE are the original EXR's.

[^Lele^]

It's not a simple change of noise threshold, no.
Next was the first release of V-Ray to break the SDK, so big the underlying changes were.

The LC has been made much better (as Vlado mentioned, which is why there is substantially less flicker in animations.), at the expense of more calculations being done by default.
There is no back-mapping these changes 1:1.
Move to the Hash map, and it's a further change.
Even one camera ray per pixel will reject some LC sample and introduce more retrace, or adopt a different BF algo to sample lighting and shaders at the given point.
To be clear, not every single sample per pixel is equal, so nope, rendering with 1SPP isn't going to unravel the problem, at all. It's just a smaller fraction of the bigger one, prone to much the same issues, if at different scales.

But many other aspects have been changed substantially since then(light, material, volume sampling. Texture filtering. Adaptive noise thresholding algorithms. the list is nigh endless. Try reading the changelogs of each SP/release since 3.7 to get an idea.), and it's impossible to naively compare: it's always going to be apple and pears, to a varying degree.

So, while the essential aspects of the rendered image are largely unchanged, they will come out cleaner, and with a much improved noise uniformity across the screen, much less flickering in animation, and a much improved light distribution in complex lighting scenarios (which will in turn require more sampling, etc etc).
This kind of statements is *always* statistical.
We're not in the business of divination, and we like to err on the cautious side of things.

As i told you in the other thread, a standard production scene will average out what calculations have been added, and what speedups have been implemented, and generally result in a quicker render.
This, before the new tech is brought to bear, which will increase speed by a wee bit still.

Notice, for clarity: when i, we at chaos say "Speed" we mean it, inherently, *to a target noise level, as measured in the output*.
We look both into the amount and into the distribution thereof.
Uniform image noise is your friend in production, and there are few engines out there able to guarantee it to the degree we do, at all, and much less so at the speeds we do it at.
A lot of effort went into it for many years (it started before v-ray 3.3), so it's not by chance.
And it's also why one can't really compare going up in versions: the new methods will change the task done, even if the settings are the same.

Lastly, "Converting to Corona" (or any other engine, for that matter) and pressing render is a meaningless endeavour, as far as benchmarking goes, so many things change between the two, in ways which are not invertible or normalisable.
Even when ensuring the closest possible parity in parameters fed to the engine, things will differ because of how each engine treats the Rendering problem.
To the point that it's difficult even for us to figure out who's quicker at a *very* specific task, at times, because of all the nuances in sampling strategies.

All this said, we do and will take any suggestion of a performance degradation *very* seriously, and will investigate it at best we can to see if it's something that we can fix.
We all take immense pride and joy in the work we do, and we wouldn't have it any other way.

[Karonte]

^Lele^ I remember you are one of the old guys ( veteran ....non for the age) that in the past tried to optimize and improve vray even with scripts ( leletools). Instead of sending you scene to check( thanks i will), can you recommend a good tutorial to try to analyze the noise level of the scenes?
How to best read the red, blue areas ... and understand whether to increase the samples or the threshold ... I have never been able to find one ,on the net, that is simple and clear. I think it is an important thing, especially for Vray users. the difference between vray and other engines is precisely the fact that you can optimize the engine ... but I have never found a clear ( simple) tutorial from chaos that would explain how.
other stupid question : Vray quick settings. I never used them in my life....Can be a dirt solution to open Vray 3.6 scene in Next4 and use this quick settings? are they balanced for the NEXT or produce the same "setting value"?
Thanks for your time.

[^Lele^]

^Lele^ I remember you are one of the old guys ( veteran ....non for the age) that in the past tried to optimize and improve vray even with scripts ( leletools). Instead of sending you scene to check( thanks i will), can you recommend a good tutorial to try to analyze the noise level of the scenes?

Post-analysis can't match what the engine does exactly: the engine considers samples, as many as needed within the pixel, while afterwards we are left only with the pixels.

A simple approach is to use one tool, whichever source it's coming from, and always only use that.
For example, one could use a noise meter from some photo retouching application (plugin or such), provided it could analyse the image as a whole and return a descriptive number.
Sticking to the same tool should -ceteris paribus- allow to get a quantitative measure which works.

A slightly more complicated approach would be to write one's own kernel, and measure noise that way.
It doesn't -for this scope- need to be fantastically accurate, or be able to detect noise from features, and so on (it can be made to do so, but with added complexity).
So, for example, you could define noise naively as the difference between a pixel, and the 8 surrounding it.
This is quite easily doable in software like Nuke (and i assume Fusion, too. i just never built one.).

How to best read the red, blue areas ... and understand whether to increase the samples or the threshold ... I have never been able to find one ,on the net, that is simple and clear. I think it is an important thing, especially for Vray users. the difference between vray and other engines is precisely the fact that you can optimize the engine ... but I have never found a clear ( simple) tutorial from chaos that would explain how.

I guess you're referring to the sampleRate RE?
If that's the case, yes, we lack a proper tutorial on sampling, and we've left it largely at this.
It's still entirely valid a logic, though.
I'll see if we can maybe pretty it up, show some better image to go with it, and clean up language and writing style.

other stupid question : Vray quick settings. I never used them in my life....Can be a dirt solution to open Vray 3.6 scene in Next4 and use this quick settings? are they balanced for the NEXT or produce the same "setting value"?

Argh, they are still around!
Nono, ignore them, they're from before 3.3 and the adaptive sampler, and we should have dropped/changed them a while back.
Thanks for mentioning those, we'll see if it's worth it to upgrade them or drop them entirely.

Thanks for your time.

You're very welcome.

[Karonte]

I followed this article and I think it can be useful to understand how to optimize. I didn't find any official Chaos description/tutorial about this optimization. Do you think is it useful? or not for Next4... They talk about "increase light/material " subdivisions but now ( in Vray next) all is automatic...

[^Lele^]

The article may have been valid at time of release (2014) but not anymore.
It's all "automatic" since v 3.3.
You simply set a max AA and a noise threshold, and let the engine take care of what needs more sampling where.
You'll notice this is essentially how every engine works (or tries to!) these days.

Don't fall in the pitfall of benchmarking this with old setups: the new will be slower because it will take care of things largely ignored before (namely: generally terrible geometric AA by virtue of too few subdivs, wildly varying noise on the image plane, and so on.).
You'll be able to make it quicker (for iteration purposes) simply adjusting the noise threshold.

[Karonte]

Ok. thanks. I will try to play with these 2 values. So..material subdivisions( 8 by default) or shadows are incremented automatically if needed? ( for example for glossy metals or if soft shadows).