Page MenuHome

Cycles GPU rendering produces a different result than CPU
Closed, DuplicatePublic

Description

System Information
Operating system: Windows-10-10.0.18362 64 Bits
Graphics card: Radeon RX 570 Series ATI Technologies Inc. 4.5.13571 Core Profile Context 19.9.3 26.20.13025.1004 (PowerColor 8GB)

Blender Version
version: 2.80 (sub 75), branch: master, commit date: 2019-07-29 14:47, hash: rBf6cb5f54494e

Short description of error
OpenCL Cycles GPU rendering produces a different result. There differences in reflections and lightness/darkness.
Some tiles seem to be correct to the CPU result.

Exact steps for others to reproduce the error
Render the attached scene with CPU, then switch render slots and render with GPU to compare difference.

I am aware of a past issue on Mac OS that appears similar, but this is on Windows.

Event Timeline

Germano Cavalcante (mano-wii) lowered the priority of this task from 90 to 50.

I can confirm.

System Information
Operating system: Windows-10-10.0.18941 64 Bits
Graphics card: Radeon (TM) RX 480 Graphics ATI Technologies Inc. 4.5.13570 Core Profile Context 19.7.2 26.20.13001.16003

Here the result using the GPU:

It appears that switching from Branched Path Tracing to Path Tracing makes the GPU render the same as the CPU render in this case. I am not sure how to confirm that for other cases.

When looking at the tile problem on the first pic (top right corner) and the one of Germano, I guess this is the same bug like the one I reported in September:
T69535: CPU + GPU render precision differences for small, far away area light

Still a really ugly one that has to be solved very soon.

I reported a bug recently that results in funky render glitches with BPT on WIN10 builds starting from 2.80 release.
Linux does not seem affected at all.
So far simply, rerendering the same image for a second time gets rid of the glitches for me.
Seems unintuitive but does the image look correct if you render it with BPT for a second time?
Also, does disabling MIS on lamps help?

I haven't tried yet if Linux is affected in the same way. Yesterday I spent hours of finding out if there is a work around for this bug, but neither BPT with different settings nor toggeling MIS does make any difference.

I rendered a demonstration video this morning for showing the effect when the area lights become smaller. Note not only the different brightness levels of the tiles (that could be avoided by disabling either CPU or GPU), but also have a look in the moire effect that can't be avoided at all:
Video: Blender 2.80 - Small area light bug

Here is the Blender file of this video. Just hit F12 to check if you can confirm it.

One small addition to my last comment:
Don't understand this bug as a problem just for area lights with sizes of millimeters or something. Depending on the szene it also can be seen using 50x50cm lights.