Manual Oversampling frustration

MattB123

Hey all

So like everyone else, I am always on a mission to find ways to speed up my renders. Im not using a bad rig, all my renders are done by GPU which for me is GTX 970 with 8GB graphics memory.

Whilst waiting for my renders I often search online for ways to speed things up and generally come across the same stuff, Some factual, and some seemingly entirely speculative.

One thing that is frustrating me though is how people talk about Manual oversampling. Every post ive read on it says with great authority "Manual oversampling works" but nobody ever provides any facts or better still numbers to show optimal oversampling percentages etc etc, they always just say "blah blah it does 1.5 times so you can do the math..." well no actually the relationship between Image resolution, and convergence required to provide a satisfactory result has not been explain in the "blah blah" so we cant do the math.

Anyway ranting aside, I want to know if anyone HAS actually run the numbers and provided a Rule of thumb for this?  I will gladly do it myself but its not a 5 minute job so I wont waste my time if someone has already done it.

agent unawares

http://buerobewegt.com/quicktip-rendering-even-faster-in-iray/

MattB123

This is one of the wishy washy posts i was talking about. Yes he recorded the times of the renders he did, but he doesn't post a conclusive set of results. 

Daywalker Designs

I suspect hard numbers aren't possible due to numerous factors, like scene differences and hardware variations.

TheKD

I have tried that a bunch of times, never got good results. Now I went back to the old fashioned way of letting the render "finish"

 

The only way I could even render at twice or more the size of the render needed, was to render it in slices. So for every image I was needing to do 6 or more render slices, assemble them together, and resize down, for dissapointing results.

I don't have a supercomputer with 8 1080ti lol. Just an average rig with a 1070 and 960

JamesJAB

So... First off, Where did you find a GTX 970 with 8GB of VRAM?  
Other than that, Iray rendering is an exercise in patience.  The GTX 980M in my laptop renders at about the same speed as your GTX 970.

I have seen zero concrete evidence that tonemapping settings have any effect on render speeds at all.

I did a test myself back in June (https://www.daz3d.com/forums/discussion/173036/dazstudio-iray-render-settings-tone-mapping-guide-with-examples#latest)
The big takeaway on my testing was that while you will still complete the same number of itterations per minute, brighter tonemapping settings will lead to a little bit higher convergence percentage per itteration.

Interior lighting is what kills Iray render speed because of the light interacting will all of the surfaces.  Iray calculates light similar to real life.

MattB123

JamesJAB said:

So... First off, Where did you find a GTX 970 with 8GB of VRAM?  
Other than that, Iray rendering is an exercise in patience.  The GTX 980M in my laptop renders at about the same speed as your GTX 970.

I have seen zero concrete evidence that tonemapping settings have any effect on render speeds at all.

I did a test myself back in June (https://www.daz3d.com/forums/discussion/173036/dazstudio-iray-render-settings-tone-mapping-guide-with-examples#latest)
The big takeaway on my testing was that while you will still complete the same number of itterations per minute, brighter tonemapping settings will lead to a little bit higher convergence percentage per itteration.

Interior lighting is what kills Iray render speed because of the light interacting will all of the surfaces.  Iray calculates light similar to real life.

Im not sure you understood my post. Manual oversampling is the practise of rendering a larger image than you need with the intention of stopping the render earlier knowing that when the image is re-sized the noise will be less apparent.

i also made a typo in my post its a 960 with 8gb not a 970

Taking on board the comments, I understand that specific times are difficult, but what should be managable is something along the lines of "assuming the final image will be 1920:1080 For a 1.5x resolution image satisfactory results can be produced at xx% convergence"

I'll have to run some experiments.