2.83 vs 2.90 CPU Benchmark

Hello world,

My first post here… Just tested 2.90 and I see an abnormally in CPU only cycle render performance. My test was using Dual Xeon E5-2620 (12 Core, 24 Thread) with 32GB ram. Using the same Hardware, I tested 2.83 & 2.90 under Windows 10 & Ubuntu 20.04.1 LTS.

Most of the test results are as expected, which 2.90 is faster (or slightly faster) than 2.83 under the same OS. However, I notice 2 tests failed my expectation… To be sure, I have ran these tests multiple times and all come out with similar numbers.

bmw27 (Windows)
2.83 - 338.471
2.90 - 373.272

koro (Windows)
2.83 - 581.214
2.90 - 996.716

Koro scene is the biggest surprise for me, it runs much slower in 2.90 (tested multiple times, on the same OS install that runs 2.83 just to be sure) I searched opendata.blender, and I found similar test results tested with AMD Ryzen 5 2600X

koro (Windows, not my test numbers, but from opendata)
2.83 - 582.596 (Slowest) 495.868 (Fastest, OC?)
2.90 - 815.995

BUT on the other hand, 2.90 also shows improvements especially under Linux

koro (Linux)
2.83 - 575.301
2.90 - 431.196

victor (Linux)
2.83 - 2039.67
2.90 - 1505.7

victor (Windows)
2.83 - 2109.77
2.90 - 1868.57

2.83 - 582.871 (Windows)
2.83 - 580.339 (Linux)
2.90 - 584.746 (Windows)
2.90 - 480.299 (Linux) :point_left:

In short, running 2.90 under Linux gives the best results for all 6 tests. And Koro scene seems to only suffer under Windows. My blender-fu is not great, what’s special in that scene that may cause the slowdown?

The slowdown in the BMW scene is a known “issue”. Here’s render times from the Blender 2.90 release notes:

“2.83” - 102 seconds
2.90 - 108 seconds

Source: https://wiki.blender.org/wiki/Reference/Release_Notes/2.90/Cycles
Original source: https://developer.blender.org/D8015

This has come about as a result of the switch to Embree as the main BVH for Cycles on the CPU in 2.90. Embree is faster in some scenes and slower in others when compared to the “old” BVH method. BMW is just one of these scenes where it is slower.

With Koro on the other hand, that is odd. The release notes show it as being faster, and as you pointed out, it is faster… on Linux.

I’m also able to reproduce a slow down with my Ryzen 9 3900X locked to 4.0Ghz:

Koro (Windows):
2.83 - 217 seconds
2.90 - 255 seconds

Koro (Linux - Debian kernel 5.7):
2.83 - 208 seconds
2.90 - 139 seconds

2.83 - https://opendata.blender.org/benchmarks/cb11543c-fe72-46fb-aaa2-2d35aa2ec973/
2.90 - https://opendata.blender.org/benchmarks/a941e3e5-35e5-457b-a136-ae0339a9623d/

2.83 - https://opendata.blender.org/benchmarks/3274cc2f-913a-44fa-ac0d-a375562a7d30/
2.90 - https://opendata.blender.org/benchmarks/f913ddb4-7996-465c-bf9b-41f3024e32aa/

As for the reason behind this descrepency, I’m not sure. I’ll leave that to someone with more experience than I have.

I saw a significant drop in render speed depending on the desktop environment. ICEWM is the fastest I’ve tested. Probably some kind a server thing without UI would be faster. Unity-Desktop not too bad. But GNOME. It’s like very bad.
You can see the height of the highlighted frame comparing to all the other graph. ( I’m using Blender Organizer to get this data ) This is I rendered one frame on Gnome. It’s more then twice the time.

So maybe including desktop environments matters.