Cycles (and sims?): Support Intel 12th Gen CPUs #99967

New Issue

H.Schoenberger · 2022-07-25T14:04:54+02:00

H.Schoenberger commented

2022-07-25 14:04:54 +02:00

System Information
Operating system: Windows 10 and 11

Blender Version
3.2.1

Short description of error
On Windows, Cycles does not use all Cores on a Intel 12th Gen CPU ("Alder Lake", contains Efficiency and Performance cores).

Cycles has to use the Windows API function SetThreadInformation().
Or Blender SetProcessInformation()

The effect is most seen if if you are rendering and minimize Blender to e.g. browse with Firefox.
Or if you use a render manager that renders in the background.

I am not sure if there are other functions in blender that might use all cores like simulations?

Test results:


Windows 10:
-	Blender UI in Foreground:         All 4 EFF cores and all 16 PERF cores.
-	Blender UI minimzed/background:   All 4 EFF cores and   0 of 16 PERF cores.
-	Console/service render            All 4 EFF cores and   0 of 16 PERF cores.

Windows 11:
-	Blender UI in Foreground:         All 4 EFF cores and all 16 PERF cores.
-	Blender UI minimzed/background:   All 4 EFF cores and   8 of 16 PERF cores.
-	Console/service render            All 4 EFF cores and   2-8 of 16 PERF cores.

**System Information** Operating system: Windows 10 and 11 **Blender Version** 3.2.1 **Short description of error** On Windows, Cycles does not use all Cores on a Intel 12th Gen CPU ("Alder Lake", contains Efficiency and Performance cores). Cycles has to use the Windows API function **SetThreadInformation()**. Or Blender **SetProcessInformation()** The effect is most seen if if you are rendering and minimize Blender to e.g. browse with Firefox. Or if you use a render manager that renders in the background. I am not sure if there are other functions in blender that might use all cores like simulations? **Test results:** ``` Windows 10: - Blender UI in Foreground: All 4 EFF cores and all 16 PERF cores. - Blender UI minimzed/background: All 4 EFF cores and 0 of 16 PERF cores. - Console/service render All 4 EFF cores and 0 of 16 PERF cores. Windows 11: - Blender UI in Foreground: All 4 EFF cores and all 16 PERF cores. - Blender UI minimzed/background: All 4 EFF cores and 8 of 16 PERF cores. - Console/service render All 4 EFF cores and 2-8 of 16 PERF cores. ```

H.Schoenberger commented

2022-07-25 14:04:54 +02:00

Added subscriber: @Schoenberger

H.Schoenberger changed title from ~~Cycles: Support Intel 12th Gen CPUs~~ to Cycles (and sims?): Support Intel 12th Gen CPUs

2022-07-25 14:13:36 +02:00

Raimund Klink commented

2022-07-25 14:55:19 +02:00

Added subscriber: @Raimund58

Sergey Sharybin commented

2022-07-25 15:39:35 +02:00

Added subscriber: @Sergey

Sergey Sharybin commented

2022-07-25 15:39:35 +02:00

As it states currently, the behavior is consistent to what is expected on Windows on a big.LITTLE architecture: processes running on a background are freeing up big cores to improve system responsivness for foreground tasks.

I am not sure forcing this behavior to change in Blender will be a good improvement. Changing process affinity/priority always gets in a way of render farms. So if anything, it should be some configuration or command line argument.

As a quick workaround you can tweak power management settings: powercfg /powerthrottling disable /path c:\path\to\blender.exe (thanks Xavier for sheding the light on this command!)

As it states currently, the behavior is consistent to what is expected on Windows on a big.LITTLE architecture: processes running on a background are freeing up big cores to improve system responsivness for foreground tasks. I am not sure forcing this behavior to change in Blender will be a good improvement. Changing process affinity/priority always gets in a way of render farms. So if anything, it should be some configuration or command line argument. As a quick workaround you can tweak power management settings: `powercfg /powerthrottling disable /path c:\path\to\blender.exe` (thanks Xavier for sheding the light on this command!)

Xavier Hallade commented

2022-07-25 16:52:41 +02:00

Added subscriber: @xavierh

Xavier Hallade commented

2022-07-25 16:52:41 +02:00

The effect is most seen if if you are rendering and minimize Blender to e.g. browse with Firefox.

The upside of this is Firefox browsing stays performant while rendering so I suppose current behavior should still be the default.
For users who would rather see background rendering finish earlier and render managers, calling powercfg /powerthrottling disable /path c:\path\to\blender.exe or rising process or threads priorities works already.

Then the open question is on if and how to expose the ability to use 100% of the resources available to users and if so, at what level.
On the code side, as you mentioned, deactivating power throttling can be done at thread level (could implement a TBB task_scheduler_observer to do that):

THREAD_POWER_THROTTLING_STATE throttling_state;
memset(&throttling_state, 0, sizeof(throttling_state));
throttling_state.Version = THREAD_POWER_THROTTLING_CURRENT_VERSION;
throttling_state.ControlMask = THREAD_POWER_THROTTLING_EXECUTION_SPEED;
throttling_state.StateMask = 0;
SetThreadInformation(thread_handle, ThreadPowerThrottling, &throttling_state, sizeof(throttling_state));

or process level:

PROCESS_POWER_THROTTLING_STATE throttling_state;
memset(&throttling_state, 0, sizeof(throttling_state));
throttling_state.Version = PROCESS_POWER_THROTTLING_CURRENT_VERSION;
throttling_state.ControlMask = PROCESS_POWER_THROTTLING_EXECUTION_SPEED;
throttling_state.StateMask = 0;
SetProcessInformation(process_handle, ProcessPowerThrottling, &throttling_state, sizeof(throttling_state));

more information on how hybrid scheduling works can be found here: https://www.intel.com/content/www/us/en/developer/articles/guide/12th-gen-intel-core-processor-gamedev-guide.html

> The effect is most seen if if you are rendering and minimize Blender to e.g. browse with Firefox. The upside of this is Firefox browsing stays performant while rendering so I suppose current behavior should still be the default. For users who would rather see background rendering finish earlier and render managers, calling `powercfg /powerthrottling disable /path c:\path\to\blender.exe` or rising process or threads priorities works already. Then the open question is on if and how to expose the ability to use 100% of the resources available to users and if so, at what level. On the code side, as you mentioned, deactivating power throttling can be done at thread level (could implement a TBB task_scheduler_observer to do that): ``` THREAD_POWER_THROTTLING_STATE throttling_state; memset(&throttling_state, 0, sizeof(throttling_state)); throttling_state.Version = THREAD_POWER_THROTTLING_CURRENT_VERSION; throttling_state.ControlMask = THREAD_POWER_THROTTLING_EXECUTION_SPEED; throttling_state.StateMask = 0; SetThreadInformation(thread_handle, ThreadPowerThrottling, &throttling_state, sizeof(throttling_state)); ``` or process level: ``` PROCESS_POWER_THROTTLING_STATE throttling_state; memset(&throttling_state, 0, sizeof(throttling_state)); throttling_state.Version = PROCESS_POWER_THROTTLING_CURRENT_VERSION; throttling_state.ControlMask = PROCESS_POWER_THROTTLING_EXECUTION_SPEED; throttling_state.StateMask = 0; SetProcessInformation(process_handle, ProcessPowerThrottling, &throttling_state, sizeof(throttling_state)); ``` more information on how hybrid scheduling works can be found here: https://www.intel.com/content/www/us/en/developer/articles/guide/12th-gen-intel-core-processor-gamedev-guide.html

H.Schoenberger commented

2022-07-25 16:59:21 +02:00

what is expected on Windows

I would state it: What is expected if you are a Windows system developer or perhaps a default windows user.
But you have to think what is expected by a Blender user. (As we got this complain from other end customers)
Other render applications have fixed it as well.

From a 3D artist view you want to render as fast as possible.
And you know if you render, it will slow down other apps on the system.
Nobody starts to play a game while he is rendering in Blender. Which is why that feature was implemented. Using foreground apps that take a lot of CPU.
The use case is completely different between a default Windows user that does not care/know about cores and CPU usage
and a 3D artist who knows about cores and cpu usage as the render time is always an issue.

You do not want that Blender uses 1/4 of your cores only, do you?
Just ask some Blender users what they think:
"Do you want that Blender always uses 1/2 or 1/4 of the cores in the case that you might want to use some other app while rendering?

If you want a reponsive UI in other apps, then there are other ways.
You have to reduce the number of render threads.
Or you can change the core affinity (which is what many render farm manager do) if someone is logged in.

Question: IN the last 20 years, have someone ever opened a feature request like these:
"Please make Blender slower, I cannot properly play a game while rendering"?
"Please let blender always use less cores by default".

Just think about it from an Blender end user perspektive.

Thanks

>what is expected on Windows I would state it: What is expected if you are a Windows system developer or perhaps a default windows user. But you have to think what is expected by a Blender user. (As we got this complain from other end customers) Other render applications have fixed it as well. From a 3D artist view you want to render as fast as possible. And you know if you render, it will slow down other apps on the system. Nobody starts to play a game while he is rendering in Blender. Which is why that feature was implemented. Using foreground apps that take a lot of CPU. The use case is completely different between a default Windows user that does not care/know about cores and CPU usage and a 3D artist who knows about cores and cpu usage as the render time is always an issue. You do not want that Blender uses 1/4 of your cores only, do you? Just ask some Blender users what they think: "Do you want that Blender always uses 1/2 or 1/4 of the cores in the case that you might want to use some other app while rendering? If you want a reponsive UI in other apps, then there are other ways. You have to reduce the number of render threads. Or you can change the core affinity (which is what many render farm manager do) if someone is logged in. Question: IN the last 20 years, have someone ever opened a feature request like these: "Please make Blender slower, I cannot properly play a game while rendering"? "Please let blender always use less cores by default". Just think about it from an Blender end user perspektive. Thanks

H.Schoenberger commented

2022-07-25 17:23:07 +02:00

Firefox browsing stays performant while rendering so I suppose current behavior should still be the default.

The issue is that Firefox does not use 10 or 16 cores.
You loose these cores. They are just idle for nothing.

Forgot to mention that you have the option to reduce the priority of the render threads to low.
This way other apps stay performant as well.

I would recommend SetProcessInformation().
As it covers all threads in Blender and batch rendering would otherwise use 4 cores of 20 only.
Which is a massive slowdown for simulations.

And for the UI mode, use an bool option right beside the thread count setting.
As the thread count is exactly what Intel tries to force.
As stated in your help about threads:
"This can be useful for example, if you want to use your computer while rendering".

>Firefox browsing stays performant while rendering so I suppose current behavior should still be the default. The issue is that Firefox does not use 10 or 16 cores. You loose these cores. They are just idle for nothing. Forgot to mention that you have the option to reduce the priority of the render threads to low. This way other apps stay performant as well. I would recommend SetProcessInformation(). As it covers all threads in Blender and batch rendering would otherwise use 4 cores of 20 only. Which is a massive slowdown for simulations. And for the UI mode, use an bool option right beside the thread count setting. As the thread count is exactly what Intel tries to force. As stated in your help about threads: "This can be useful for example, if you want to use your computer while rendering".

Germano Cavalcante commented

2022-07-26 16:14:28 +02:00

Changed status from 'Needs Triage' to: 'Needs Developer To Reproduce'

Germano Cavalcante commented

2022-07-26 16:14:28 +02:00

Added subscriber: @mano-wii

Jesse Yurkovich commented

2022-08-05 00:28:44 +02:00

Added subscriber: @deadpin

Jesse Yurkovich commented

2022-08-05 00:28:44 +02:00

I did a bit digging into the OpenData benchmark and the 12th gen's do seem to exhibit much wider variation compared to 10th on all scenes and maybe this is partly to blame?
When downloading the raw json dump and looking at some basic stats for "Intel Core i7-10700 CPU @ 2.90GHz" vs "12th Gen Intel Core i7-12700" on version 3.1.0:

scene	10th (min, max, std. dev)	12th (min, max, std. dev)
monster	63, 93, 10.7	25, 157, 24.3
junkshop	0.8, 56, 12	1.8, 84, 19.33
classroom	28, 43, 5.9	11, 74, 12.9

Doesn't the behavior here also change if the user is using a multi-monitor setup? I remember reading that this type of balancing will be disabled in that case since it's not clear if 1 app is in the background vs. not etc.

We should probably just use all the threads and let the user pick how many, if they don't want All, with the existing option that is available in the UI.

I did a bit digging into the OpenData benchmark and the 12th gen's do seem to exhibit much wider variation compared to 10th on all scenes and maybe this is partly to blame? When downloading the raw json dump and looking at some basic stats for "Intel Core i7-10700 CPU @ 2.90GHz" vs "12th Gen Intel Core i7-12700" on version 3.1.0: | scene | 10th (min, max, std. dev) | 12th (min, max, std. dev) | -- | -- | -- | | monster | 63, 93, 10.7 | 25, 157, 24.3 | junkshop | 0.8, 56, 12 | 1.8, 84, 19.33 | classroom | 28, 43, 5.9 | 11, 74, 12.9 Doesn't the behavior here also change if the user is using a multi-monitor setup? I remember reading that this type of balancing will be disabled in that case since it's not clear if 1 app is in the background vs. not etc. We should probably just use all the threads and let the user pick how many, if they don't want All, with the existing option that is available in the UI.

Philipp Oeser removed the

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Cycles (and sims?): Support Intel 12th Gen CPUs #99967