Author Topic: Threadripper vs Intel Xeon: Different render outputs in noise level  (Read 3520 times)

2017-09-02, 10:16:49

johan belmans

  • Primary Certified Instructor
  • Active Users
  • ***
  • Posts: 761
    • View Profile
    • belly.be
Hi there,

At the moment our renderfarm consists out of Intel Xeon machines, we are planning to buy some AMD Threadripper machines to enlarge our farm.
I was told that the Threadripper is handling floating points in a different way then Intel which could result in different noise level outputs for the same render.
Has anyone experienced this? Or other issues? Or no issues?
« Last Edit: 2017-09-04, 09:21:59 by johan belmans »

2017-09-03, 18:30:50
Reply #1

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
it could probably lead to different random seed, but definitely not to different noise amount.
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2017-09-04, 09:23:42
Reply #2

johan belmans

  • Primary Certified Instructor
  • Active Users
  • ***
  • Posts: 761
    • View Profile
    • belly.be

2017-09-04, 10:03:12
Reply #3

Ryuu

  • Former Corona Team Member
  • Active Users
  • **
  • Posts: 654
  • Michal
    • View Profile
Threadrippers are basically two Ryzen chips in a single package. We didn't do any extensive testing on Ryzens ourselves, but so far we haven't seen any difference. I doubt that there would be any difference since both AMD and Intel always tried to make their CPUs compatible.

2017-09-04, 11:13:39
Reply #4

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
Also CPUs are more deterministic and exact than people generally think. There is no quality difference in the results of operations, the ones and zeroes produced are not structured in better way or more artisanal in one or the other architecture. Even the famous pentium FDIV bug that everyone was making fun out of was a mistake in order of 0.007%, triggered once per 9 000 000 000 calculations.
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2017-09-04, 11:23:45
Reply #5

Ryuu

  • Former Corona Team Member
  • Active Users
  • **
  • Posts: 654
  • Michal
    • View Profile
Well, the physical CPUs are also subject to various non-deterministic manufacturing variations. As long as you stay within the recommended frequency/voltage/temperature limits, everything is fine. Once you are past these limits, nothing is certain.

With all the overclocking going on with Ryzen/Threadripper CPUs, I wouldn't be surprised if some "lucky" guy pushed his CPU just slightly over its physical limits, where the system is still stable most of the time, but the floating point calculations sometimes do give incorrect results.

2017-09-04, 14:37:24
Reply #6

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
well if that ever happened to actually be visible in the final image, it must have been millions of times (each pass and each pixel is independent), and there is strong possibitlity he would just switched a bit in pointer and got segfault along the way ;)
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)