Cycles perf regression in Blender 4.3.0 Release vs Release Candidate #130580

New Issue

blu42 · 2024-11-20T08:46:15+01:00

blu42 commented

2024-11-20 08:46:15 +01:00

System Information
Operating system: Windows 11 Home 24H2 build 26100.2314
Graphics card: none; CPU Snapdragon X1E-78-100

Blender Version
Broken: 4.3.0. release, 2b18cad88b13, 2024-11-19
Worked: 4.3.0 release candidate d3d26233fcb8; 2024-11-15

Performance regression in scene bmw27_cpu
Regression of approx. 10 seconds in scene bmw27_cpu
Release candidate: ~2m20s
Release: ~2m30s

Exact steps for others to reproduce the error
Render in Cycles CPU renderer.

**System Information** Operating system: Windows 11 Home 24H2 build 26100.2314 Graphics card: none; CPU Snapdragon X1E-78-100 **Blender Version** Broken: 4.3.0. release, `2b18cad88b13`, 2024-11-19 Worked: 4.3.0 release candidate `d3d26233fcb8`; 2024-11-15 **Performance regression in scene bmw27_cpu** Regression of approx. 10 seconds in scene bmw27_cpu Release candidate: ~2m20s Release: ~2m30s **Exact steps for others to reproduce the error** Render in Cycles CPU renderer.

blu42 added the

labels 2024-11-20 08:46:15 +01:00

YimingWu commented

2024-11-20 14:35:33 +01:00

Hi! I'm not sure if this would be a significant regression, 10s in 2min20s is like ~7%, and this kind of fluctuation could happen based on different system load.

Just to make sure, you are using the native ARM build? If you are using the x86_64 one then your CPU transplier could greatly affect the performance.

Hi! I'm not sure if this would be a significant regression, 10s in 2min20s is like ~7%, and this kind of fluctuation could happen based on different system load. Just to make sure, you are using the native ARM build? If you are using the x86_64 one then your CPU transplier could greatly affect the performance.

blu42 commented

2024-11-20 15:00:29 +01:00

hi @ChengduLittleA

Non-native code was not involved in these results -- both versions were arm64. Release Candidate was acquired from earlier daily builds, Release version was the .MSI bundle, correctly identified by the browser download dialogue as 'windows arm64 300MB'

The machine all tests were run on was a Lenovo Yoga 7x with the specified Snapdragon chip. Win11 System Performance mode was set to 'High performance', repeated runs were performed under idle machine conditions. I don't have the logs but I can re-run the tests if full logs are needed.

hi @ChengduLittleA Non-native code was not involved in these results -- both versions were arm64. Release Candidate was acquired from earlier daily builds, Release version was the .MSI bundle, correctly identified by the browser download dialogue as 'windows arm64 300MB' The machine all tests were run on was a Lenovo Yoga 7x with the specified Snapdragon chip. Win11 System Performance mode was set to 'High performance', repeated runs were performed under idle machine conditions. I don't have the logs but I can re-run the tests if full logs are needed.

Alaska commented

2024-11-20 15:35:03 +01:00

Running the full Cycles benchmark suite here with a AMD Ryzen 9 3900X on Windows, there is a small performance difference in some scenes (In the worst case the final release of Blender 4.3 takes 2% longer to render, but most of the time it's less than 1%). I would class these as "within standard fluctuation of testing" and wouldn't say my results are a performance regression.

In Blender chat Yimming Wu asked if there were any changes to Cycles in that time period that might of caused this issue, and there's only one (1a40efbded), but I'm highly doubtful it would make renders 7% longer on this platform.

@blu42 are you able to run Blender with extra logging for us?

Open command prompt
Run /path/to/blender --debug-cycles (Use the final release here)
Open the scene you're using for testing and turn off adaptive sampling if it isn't already off (BMW should have it off)
Render the scene
In the command prompt you will see something like this at the end. Share it with us.

Full path tracing report

Path tracing on: AMD Ryzen 9 3900X 12-Core Processor (CPU) (24 threads) [CPU]

Render Scheduler Summary

Mode: Background
Resolution: 1920x1080

Adaptive sampling:
  Use: False

Denoiser:
  Use: False

Time (in seconds):
                                       Wall              Average
          Path Tracing           258.487731             0.063107
        Display Update             1.096877             0.008438

  Total: 259.584609

Rendered 4096 samples in 259.623028 seconds
Average time per sample: 0.063385 seconds
I1121 03:33:07.627838 11904 session.cpp:479] Total render time: 259.642
I1121 03:33:07.628867 11904 session.cpp:480] Render time (without synchronization): 259.623

Do the same for the release candidate

This helps us see where the performance may be lost.

Running the full Cycles benchmark suite here with a AMD Ryzen 9 3900X on Windows, there is a small performance difference in some scenes (In the worst case the final release of Blender 4.3 takes 2% longer to render, but most of the time it's less than 1%). I would class these as "within standard fluctuation of testing" and wouldn't say my results are a performance regression. In Blender chat Yimming Wu asked if there were any changes to Cycles in that time period that might of caused this issue, and there's only one (1a40efbdedef3fbfeaa776f7e1ab37ccbc9d0421), but I'm highly doubtful it would make renders 7% longer on this platform. @blu42 are you able to run Blender with extra logging for us? - Open command prompt - Run `/path/to/blender --debug-cycles` (Use the final release here) - Open the scene you're using for testing and turn off adaptive sampling if it isn't already off (BMW should have it off) - Render the scene - In the command prompt you will see something like this at the end. Share it with us. ``` Full path tracing report Path tracing on: AMD Ryzen 9 3900X 12-Core Processor (CPU) (24 threads) [CPU] Render Scheduler Summary Mode: Background Resolution: 1920x1080 Adaptive sampling: Use: False Denoiser: Use: False Time (in seconds): Wall Average Path Tracing 258.487731 0.063107 Display Update 1.096877 0.008438 Total: 259.584609 Rendered 4096 samples in 259.623028 seconds Average time per sample: 0.063385 seconds I1121 03:33:07.627838 11904 session.cpp:479] Total render time: 259.642 I1121 03:33:07.628867 11904 session.cpp:480] Render time (without synchronization): 259.623 ``` - Do the same for the release candidate This helps us see where the performance may be lost.

Alaska added

Status

Needs Information from User

and removed

Status

Needs Triage

labels 2024-11-20 15:35:16 +01:00

Lukas Stockner commented

2024-11-20 15:52:37 +01:00

For even more info, use /path/to/blender --debug-cycles -- --cycles-print-stats (yes, the extra -- is intentional).

For even more info, use `/path/to/blender --debug-cycles -- --cycles-print-stats` (yes, the extra -- is intentional).

Anthony Roberts commented

2024-11-20 18:12:36 +01:00

I'm only seeing a ~2-3s render time change on my Thinkpad T14s between the RC and Release versions, which I would also call statistical noise.

I'll give it a try on an 8cx Gen3, where the effect could well be more pronounced

I'm only seeing a ~2-3s render time change on my Thinkpad T14s between the RC and Release versions, which I would also call statistical noise. I'll give it a try on an 8cx Gen3, where the effect could well be more pronounced

Anthony Roberts commented

2024-11-20 18:56:03 +01:00

Sadly, I see no difference on one of those either.

8cx Gen3 Release version: 5:49.47 (average over 2)
8cx Gen3 RC Version: 5:48.86 (average over 2)

Are you always testing one after the other? Could it maybe be laptop thermals kicking in and throttling the CPU back slightly at the end of the second test?

Sadly, I see no difference on one of those either. 8cx Gen3 Release version: `5:49.47` (average over 2) 8cx Gen3 RC Version: `5:48.86` (average over 2) Are you always testing one after the other? Could it maybe be laptop thermals kicking in and throttling the CPU back slightly at the end of the second test?

👀 1

blu42 commented

2024-11-20 19:09:46 +01:00

Here are the logs from --debug-cycles. Trying to collect --cycles-print-stats as well skews the times too much on this machine.

Here are the logs from `--debug-cycles`. Trying to collect `--cycles-print-stats` as well skews the times too much on this machine.

Blender 4.3 Release Candidate.txt

14 KiB

Blender 4.3 Release.txt

14 KiB

blu42 commented

2024-11-20 19:14:14 +01:00

@Anthony-Roberts

Sadly, I see no difference on one of those either.

8cx Gen3 Release version: 5:49.47 (average over 2)
8cx Gen3 RC Version: 5:48.86 (average over 2)

Are you always testing one after the other? Could it maybe be laptop thermals kicking in and throttling the CPU back slightly at the end of the second test?

I always give the machine a minute to settle in-between runs. Sequential same-version runs produced in this fashion give consistent times. Also laptop is set to High performance power profile and the fans quickly go full tilt.

@Anthony-Roberts > Sadly, I see no difference on one of those either. > > 8cx Gen3 Release version: `5:49.47` (average over 2) > 8cx Gen3 RC Version: `5:48.86` (average over 2) > > Are you always testing one after the other? Could it maybe be laptop thermals kicking in and throttling the CPU back slightly at the end of the second test? I always give the machine a minute to settle in-between runs. Sequential same-version runs produced in this fashion give consistent times. Also laptop is set to `High performance` power profile and the fans quickly go full tilt.

Alaska commented

2024-11-21 03:25:12 +01:00

Trying to collect --cycles-print-stats as well skews the times too much on this machine.

Do the results still end up with teh final release being noticably slower? If so the information is still useful to us.

Also thanks for the logs, they confirm that the differen in time comes from the rendering. And work seems to be scheduled identically so the cause is likely the rendering code not some other random variable.

> Trying to collect `--cycles-print-stats` as well skews the times too much on this machine. Do the results still end up with teh final release being noticably slower? If so the information is still useful to us. Also thanks for the logs, they confirm that the differen in time comes from the rendering. And work seems to be scheduled identically so the cause is likely the rendering code not some other random variable.

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Cycles perf regression in Blender 4.3.0 Release vs Release Candidate #130580