Which computer advice?

[Digital Magic]

Wow that is surprising and concerning. I've ordered a new machine (threadripper 3970x) specifically to speed up super long renders I am having with V-ray 3.7. I could have bought an i7 or i9 for considerably less.
It's a lot of money for me but I decided it was worth it because I can't wait 2 weeks for a scene to render on my old i7. Soon I'll be using Phoenix also.
Now I'm wondering if I made a costly mistake.

[I-REEL]

Hi Digital Magic, There are perf issues with AMD Epyc / Vray, On the other side Threadripper is performing quite well with Vray (even if it should perform even better when looking at specs), it is still best performance / cost vs xeon in my opinion). You will definitely see a huge improvements in rendering time. Don't worry.
(We have both Epyc and threadripper config here.)
Small advice : you should always check Vray benchmark results before choosing a CPU. If you look at it you will see that 3970X is a good choice
Keep us posted with your results.

[Digital Magic]

Hi Digital Magic, There are perf issues with AMD Epyc / Vray, On the other side Threadripper is performing quite well with Vray (even if it should perform even better when looking at specs), it is still best performance / cost vs xeon in my opinion). You will definitely see a huge improvements in rendering time. Don't worry.
(We have both Epyc and threadripper config here.)
Small advice : you should always check Vray benchmark results before choosing a CPU. If you look at it you will see that 3970X is a good choice
Keep us posted with your results.

Hi I-REEL. Thank you much. I looked at lots of benchmarks and the threadrippers seemed like a good choice based on the charts and what people were generally saying. That's why I made the jump. Then started questioning it after reading some issues.
Thanks again for sharing. Hopefully it will work out ok

[matanov]

The poor performance of the AMD vs that Intel is very likely to be because of the MSR. We will hide this option in the future releases because people seem to abuse it without knowing what it actually does.

We'll check this particular case of course, if you provide the scene.

[seandunderdale]

The poor performance of the AMD vs that Intel is very likely to be because of the MSR. We will hide this option in the future releases because people seem to abuse it without knowing what it actually does.

We'll check this particular case of course, if you provide the scene.

Can you quickly give a run down of its use? From what I understand it biases the renderer to either glossy sampling (reflections, DOF) or to AA. But to me its a weird either / or. Also, there is a lower limit of 0 but no upper limit. What would be ridiculous settings?

[^Lele^]

Can you quickly give a run down of its use? From what I understand it biases the renderer to either glossy sampling (reflections, DOF) or to AA. But to me its a weird either / or. Also, there is a lower limit of 0 but no upper limit. What would be ridiculous settings?

I shall try and explain this.
Know the real way V-Ray works is a bit more nuanced than i'll make it out to be here, however the main concepts will hold.
Also, you can consider this as official: it's not my personal opinion, rather a report.

To accelerate shading, for each Camera ray shot, V-Ray (CPU!) will shoot a number of secondary rays.
In the old days, those were determined by the subdivs on shaders and lights.
AA was always at a premium (16 rays for AA seemed like a lot, back in the early '00s), and so one did most of the evaluations with secondary, specific rays.
The old balance would be 1 to 64 (with the default of 8 subdivs) or even higher (i have seen stuff with 256 SUBDIVS in production).
This would lead, invariably, to over/under sampling, flickering geometric and texture detail, uneven noise levels across an image and a sequence, and endless debugging of scenes to figure out why a plane and sphere would take a week to converge.

Enter V-Ray 3.3 and the new adaptive sampler.
With it, local subdivs are ignored, in favour of an analysis of the noise levels in samples, and automatic adaptation of sampling based on the actual contents of the image.
What changed, compared to old setup styles, is that now AA is free to reach the levels it deems right for the image at hand and the set noise threshold (f.e. thin geo or texture detail, Dof and moblur of variable intensity/length, where low AA would produce flickering or excessive noise.), and as such, it will generally reach much higher numbers than 16 Camera rays.

So, with local subdivisions ignored, MSR takes center stage: the default value of 6 will spawn 6 secondary rays (of any of the needed types.) per AA sample.
The engine is also smart, in that it will change this balance to optimise sampling of stuff like Hair and Fur, or in the case of DoF and MoBlur (where Camera rays are a must.).
Because it's not forced by the user, but driven by the sampling needs alone, the engine will ensure (near) optimal resource ustilisation to reach the set noise threshold.

There are generally two (and a half!) main concerns by users about this approach:

a1) Why so few? I used to use 64 subdivs everywhere!
While MSR seems a tiny number, compared to the huge counts of local subdivs of old, it's just an impression: with AA able to reach, by default, 10000 rays, the pixels which will need it (and only those!) will be able to hit 60000 secondary samples per pixel (plos 10k AA. 70k total), which far exceed all but the most extreme of setups of old (f.e. 64 subdivs and 4 AA would make 4096*16 samples, or 65536+16 total. Still less.).
a2) I know better: this has to be much higher!
Only in exceedingly simple cases (what i call the plain wall syndrome): we took extreme care testing this out before settling for a default (feel free to reread the whole thread, instead of just vlado's post.).
What looks like a speedup for the plain areas of a scene will result in a slowdown whenever AA will have to work harder to ensure thin details (or, currently, overbrights) are properly sampled through Camera (AA) rays.

b) I hate Universal Settings! It's so Slow!
It was before V-Ray 3 SP3.
And it was because the "MSR" back then was the number of subdivs per material and light.
So, by default, it'd have been the 1:64 ratio of above.
And it'd have still produced uneven noise across the image or sequence, if less than with low max AA.

This is as concise a rundown to MSR, and its history, as i can give you.
If you felt like reading more on the subject, and on other people's concerns about it, you can read here, and here, and here.

TL;DR remains "forget about it.". ^^

[Vipera]

I have created a simplified version of the scene and made more renders on AMD and Intel computers with different settings. You could download scene and rendered samples from this link.
https://www.dropbox.com/sh/at3d5k7su...ebA3jp31a?dl=0

My findings from tests on Intel i7-6950X 10-core and AMD Ryzen Threadripper 3960X 24-Core CPUs:
1. MSR doesn't play a role in Progressive render as apposite to Bucket render where it helps to remove noise without significantly increasing render time on Intel computer. But AMD has a big slowdown in render time with increasing MSR.
2. Bucket render is twice slower on AMD computer.

I have a question. Why Vray benchmark shows that Ryzen 3960X is faster than i7-6950X but test renders do not prove this?

[^Lele^]

I have a question. Why Vray benchmark shows that Ryzen 3960X is faster than i7-6950X but test renders do not prove this?

Because the benchmark doesn't allow users to tweak settings?

[dlparisi]

Because the benchmark doesn't allow users to tweak settings?

But looking at his test renders it seems that renders using the default 6 MSR he gets times of 1:06 on the 6950x and 2:02 on the 3960x. Granted he still is using max subdivs of 4 on both rather than the default 24 but the times don't correlate with the benchmark results at all.

[Alex_M]

I have created a simplified version of the scene and made more renders on AMD and Intel computers with different settings. You could download scene and rendered samples from this link.
https://www.dropbox.com/sh/at3d5k7su...ebA3jp31a?dl=0

My findings from tests on Intel i7-6950X 10-core and AMD Ryzen Threadripper 3960X 24-Core CPUs:
1. MSR doesn't play a role in Progressive render as apposite to Bucket render where it helps to remove noise without significantly increasing render time on Intel computer. But AMD has a big slowdown in render time with increasing MSR.
2. Bucket render is twice slower on AMD computer.

I have a question. Why Vray benchmark shows that Ryzen 3960X is faster than i7-6950X but test renders do not prove this?

Hey Vipera ,

I was curious about your issue with Threadripper 3960x since I own its "bigger brother", the 32-core 3970x. So I've just downloaded your scene to check out. To simulate the 24-core 3960x I disabled 8 cores in BIOS to effectively make my 3970x into a 24-core 3960x.

Before rendering, I made sure that displacement was enabled, sampler was set to Bucket and that MSR is set to 32 which according to you is slow on 3960x. The results are pretty interesting (screenshots down below) - rendering took exactly 1 minute at 1920x1080 and 3:52 min at 3840x2160. Not even remotely as slow as your results with the 3960x which for some reason takes 70 min at 3840x2160 or 17.5 times slower.

So the good news is, as I thought, AMD Threadripper 3000 CPUs aren't slower than a 4 year old Intel CPU with 2.4 times less cores. In fact, they are 6 times faster than Intel i7 6950x. Bad news is, you'll have to check what in your system causes render times to be that high because, as I showed, my simulated 24-core Threadripper 3960x is 17.5 times faster than your Threadripper 3960x.

As promised, screenshots:

[super gnu]

unless im missing something your render lacks the foam that is present in the original. this could have an effect on rendertimes no?

edit: following this thread closely as i have the parts for a 3970x workstation on their way to me.

[Vipera]

So the good news is, as I thought, AMD Threadripper 3000 CPUs aren't slower than a 4 year old Intel CPU with 2.4 times less cores. In fact, they are 6 times faster than Intel i7 6950x. Bad news is, you'll have to check what in your system causes render times to be that high because, as I showed, my simulated 24-core Threadripper 3960x is 17.5 times faster than your Threadripper 3960x.

Alex_M, thanks a lot for making tests! At least I know it works for someone. Could you make the render test with all your cores, please?

[Vipera]

unless im missing something your render lacks the foam that is present in the original. this could have an effect on rendertimes no?

I simplified this scene from original to make the scene smaller for download. I made render tests with this simplified scene in 1920x1080 resolution. With the render settings what Alex_M tried I have render time 2m51s to compare with his 1m2s. This needs to be investigated.

[dlparisi]

Keep in mind too though that Vipera is running on 3.6 and Alex is on 4.3 which could affect times.

Further adding to the confusion though (at least for me), is that when I render on my system (a dual Xeon 2696v3 w/ 36 cores, 72 threads, scoring 23652 on the VRay Next Benchmark) I get times of 2m15s for 32MSR and 1m30s for 6MSR. Comparing these to Vipera's times on the 6950x, my 32MSR is somewhat faster then his (2m15s vs. 2m51s) but somewhat slower than his for the 6MSR (1m30s vs. 1m5.6s). There is the 5-10 seconds at the start for the displacement that is single threaded which could skew things a bit but it seems odd given that the 6950x has a benchmark of 12796 and my xeon is at 23652. It seems like the xeons should easily beat his 6950x all around (nevermind about his really odd TR times). I'm also on 4.3.1.

What's going on?

[Alex_M]

Keep in mind too though that Vipera is running on 3.6 and Alex is on 4.3 which could affect times.

He said that he tested with Vray Next (version 4.xx) too. His quote:

...But when I render in 3ds Max using V-Ray 3.60 or V-Ray Next my i7 renders faster than Ryzen.

Alex_M, thanks a lot for making tests! At least I know it works for someone. Could you make the render test with all your cores, please?

Sure. Please see below. This is stock (no overclocking), all 32 cores enabled.