AMD vs INTEL Far Bump Calculation

[benjurio]

Dear Corona Team,

We've encountered a strange issue which might be related to CPU manufacturers. In our office, we have dozens of Intel Xeon and i7 machines, as well as AMD Ryzen machines.
We are working on an animation, which involves river water rendering. The river water geometry is a simple plane, with a simple material, with bump, diffuse, etc, so no simulation included.

We experience that our ryzen and intel machines render the water differently, but the difference only occurs in the far distant areas, close to the horizon. I've attached the same frame rendered by a ryzen and an intel xeon machine. (We have the exact the same version of corona, max, plugins, scripts installed on all machines.)

We tried to investigate the issue by turning off diffuse, reflection, refraction, bump, etc, and what we experience is that this problem is most likely related to bump. If we turn off bump, the renders are the same. We turn bump on, it's different. Changing the bump texture in any way does not help.

Do you happen to know the reason or the solution for this issue?

Thank you,
Bence | ZOA Studio

[rowmanns]

Hi,

Thanks for reporting this. We are aware of this issue and we are looking into it.

I will do some further investigations and get back to you when I have some more information.

Thanks,

(Internal ID=323484663)

[romullus]

Duplicate: https://forum.corona-renderer.com/index.php?topic=27003.0

[Juraj]

This is a real thing :- O ? Wow.

Interesting I've never noticed since I have a lot of nodes on both AMD and Intel. This could be nasty in animation no?

[Fluss]

Yep, I'm worried as well and might reconsider buying threadrippers anytime soon, just for consistency purposes. I don't say AMD chips are the cause of the issue. Actually, it looks like Intel's renders have artifacts that AMD's ones do not have.

Are you able to upload that scene so we can test it on our side?

[benjurio]

Hello Everyone,

Corona Team, thank you for investigating this.

I've attached the max, but this bug can simply be recreated by adding a simple bumpy material to a plane, and look at it from a shallow angle.
I've also attached 2 other testrenders, where the difference is more visible. The issue occurs with really high glossiness (0,9+), and huge bumps - that's why it's mostly obvious with open water scenes.

Hope this helps with the troubleshooting.

Thanks,
Bence

[romullus]

WOW, the difference is massive and to me it looks that amd output is correct and xeons shows some serious artefacts. I wonder if it affects all intel CPUs or xeons only. I would like to test the scene on i7, but i can't open it. Would it be possible to resave in max 2016 format?

[benjurio]

Yes, AMD seems to be the proper one. Btw, not only xeons, but our older i7 cpu-s produce the same wrong image as well.

I've attached the 2016 version max file.

[Fluss]

Just to let you know and to avoid any confusion, I'm not part of the corona team.

I tested on both xeon V3 and V4 and I can confirm that I have the same artifacts as you. Really strange.

[romullus]

Thank you for the scene! I can confirm, on i7 it renders exactly like in xeon.jpg example.

[agentdark45]

As an aside to this, while the devs are looking at bump mapping - would it also be possible to look into how micro bump details are handled?

Myself and others on the forum have mentioned that when it comes to micro details (dense micro metal scratches/bumps for example), the result is often view and resolution dependent which can wreak havoc when going from a low res test render to a high res production render.

A quick example of this would be to add a stretched noise bump map to simulate anistropy/realistic reflection dulling on a metal pot. I can do this easily in Fstorm with just bump mapping (sorry to beat the Fstorm drum again), but I can't with Corona. Please see the video below for an example of this in action:

[Juraj]

Is the F-Storm interpolated that sharply specifically in bump or in general? Yeah if Bump/Normal would drive roughness at resolution-independent way, that would be...kick ass.

[agentdark45]

Is the F-Storm interpolated that sharply specifically in bump or in general? Yeah if Bump/Normal would drive roughness at resolution-independent way, that would be...kick ass.

I think it's an engine wide approach to texture/AA filtering as bitmaps in general render with more clarity at distance, Dubcat touched on this in his thread. Whatever it is, it's goddamn magical - the smudge/dust maps on the metal pot and cooker surface are so well sampled and devoid of any "cgi" sharpness, whilst also retaining amazing detail.

[Juraj]

Yeah better texture filtering/interpolation would be good in general. It's one of the reason I render so high-res always so I get as much detail possible before downsampling. And still everything looks too smooth to me...

[Njen]

At Siggraph 2018 I attended a Pixar talk where they introduced in Renderman their solution to what they called 'bump to roughness' when rendering the fine detail of cloth at different distances to camera. The results were night and day.