The importance of "binding" CPU interrupts for a device to a specific CPU core to reduce overall CPU temperatures.

[mql]

I want to hear opinions and recommendations about: "The importance of "binding" CPU interrupts for a device to a specific CPU core to reduce overall CPU temperatures."

For example, I can give my situation when one of the processor cores is thermally throttling more than the others. Therefore, I am interested in the possibilities of how to reduce the load on specific processor cores.

I tried to solve the mentioned problem using the "Microsoft Interrupt Affinity Tool" by changing the cores available for VIDEO, AUDIO and WiFi devices, selecting those CPU cores that show the lowest temperature during their operation.

 

[Night]

It depends on the application you're running, and if the certain workload can be parallelized across the CPU's cores (multithreading). If not, next available less-busy CPU core will be used for non-parallelized workload therefore increasing that same core temperature.

 

[mql]

It depends on the application you're running, and if the certain workload can be parallelized across the CPU's cores (multithreading). If not, next available less-busy CPU core will be used for non-parallelized workload therefore increasing that same core temperature.

I have been experimenting for several months with selecting specific kernels for Video and Audio devices. So far I believe this has had a positive effect, reducing the load (temperatures) on the first CPU core and possibly another one of the two fastest CPU cores.
However, there are applications that constantly load the fastest cores in the CPU, causing those cores to directly overheat and throttle the CPU.

I'm interested in removing background processes from the fastest CPU cores. But I don't know how to find out which devices are using certain CPU cores in the background.

 

[dgianstefani]

Literally just use process lasso.

 

[mql]

Literally just use process lasso.

But this program uses a lot of CPU resources while running.

+
I think this will be a good program. Thank you.

 

[Assimilator]

The smartest and easiest way to solve thermal throttling is by better cooling. This is the exact opposite of smartest.

 

[mql]

The smartest and easiest way to solve thermal throttling is by better cooling. This is the exact opposite of smartest.

In terms of cooling, almost everything possible has already been done on this computer.

A month has passed since the thermal paste was applied to the processor chip using circuit boards that hold the thermal paste.

At the moment this is the best solution I have. TS test result with 7680M.
View attachment 323621

120M / 5.937 = 20.21
7680M / 373.105 = 20.58

 

[Night]

I'm interested in removing background processes from the fastest CPU cores. But I don't know how to find out which devices are using certain CPU cores in the background.

OS scheduler gives work to the CPU, which it will complete in non-linear way also known as Out of Order Exececution, which means that CPU probably has more work on 1 core than you think, so I think there's no way to regulate what kind of work and of how many programs will be done on each core. That is up to the CPU and its architecture.

 

[Onasi]

I think the question that no one asked yet is twofold - what are the specs that you are working with and what workload are you applying that causes the CPU to throttle in the first place? I doubt this is idle temps, maybe you are hitting the cores with a heavy AVX workload in which case the whole discussion is moot.
Anywho, trying to wrangle background processes to a certain core while your system runs into TJmax is one of the weirdest “cart before horse” cases I’ve seen in a while.

 

[mql]

I think the question that no one asked yet is twofold - what are the specs that you are working with and what workload are you applying that causes the CPU to throttle in the first place? I doubt this is idle temps, maybe you are hitting the cores with a heavy AVX workload in which case the whole discussion is moot.
Anywho, trying to wrangle background processes to a certain core while your system runs into TJmax is one of the weirdest “cart before horse” cases I’ve seen in a while.

OS scheduler gives work to the CPU, which it will complete in non-linear way also known as Out of Order Exececution, which means that CPU probably has more work on 1 core than you think, so I think there's no way to regulate what kind of work and of how many programs will be done on each core. That is up to the CPU and its architecture.

Perhaps I'm thinking wrong. When I started researching what CPU thermal throttling is doing, I noticed that depending on which CPU cores are connected to the video device (NVIDIA Graphics), the CPU cores that start thermal throttling change depending on which cores are selected. This is how the idea came about that there are processes that constantly load up certain CPU cores, and when other programs also start using the same cores heavily, the CPU overheats.

 

[Onasi]

You ARE thinking wrong. Device driver interrupts are an insignificant load on any reasonably modern CPU. They DEFINITELY should not cause the chip to throttle, whatever other load is being run is irrelevant. Otherwise, everyones PC would be throttling like mad, which is obviously not the case.
Either somehow something on the software side is FUBAR on your system, your cooling is straight up insufficient or for some reason your CPU runs significantly out of spec. None of these potential issues are fixable via puttering around with affinities. You are trying to treat sneezing while suffering from smallpox.

 

[mql]

You ARE thinking wrong. Device driver interrupts are an insignificant load on any reasonably modern CPU. They DEFINITELY should not cause the chip to throttle, whatever other load is being run is irrelevant. Otherwise, everyones PC would be throttling like mad, which is obviously not the case.
Either somehow something on the software side is FUBAR on your system, your cooling is straight up insufficient or for some reason your CPU runs significantly out of spec. None of these potential issues are fixable via puttering around with affinities. You are trying to treat sneezing while suffering from smallpox.

In that case I don't understand it:

 

[Onasi]

This all refers to latency mostly, not load and definitely not temps. It’s also a bit of a meme since the tool is mostly useful for testing or enterprise workloads and the whole “dedicating a core to GPU is totally going to make your K/D in CS higher” is something that became retroactively popularized among input lag nutters who swear up and down that everything, moon phases included, affects the nebulous “mouse feel”.
But that particular tangent is not in scope of this thread.

Edit: You still have provided no specs and no specifics on your workloads. I am not a fan of fumbling in the dark and neither are most people here. So far all we are doing are discussing theory with no tangible topic.

 

[mql]

This all refers to latency mostly, not load and definitely not temps. It’s also a bit of a meme since the tool is mostly useful for testing or enterprise workloads and the whole “dedicating a core to GPU is totally going to make your K/D in CS higher” is something that became retroactively popularized among input lag nutters who swear up and down that everything, moon phases included, affects the nebulous “mouse feel”.
But that particular tangent is not in scope of this thread.

Of course, I’m not “swear” here that in my case this is exactly the case, that with the help of the “Microsoft Interrupt Affinity Tool” the load was removed from specific processor cores.

Edit: You still have provided no specs and no specifics on your workloads. I am not a fan of fumbling in the dark and neither are most people here. So far all we are doing are discussing theory with no tangible topic.

Everything has already been written in this thread once
CPU i7-12700H, RING EDP OTHER RED, DELL G15 5520, motherboard 05J7DY. | TechPowerUp Forums

 

[Onasi]

…you do realize that you could have just provided a link to that thread from the start? We’re not clairvoyant.
Anyway, so it’s a laptop. This changes/complicates things. From what’s been done and discussed in that thread, it seems the consensus was that both the cooling is not great and that the CPU as configured by manufacturer was running into issues. It’s hard to definitively say whether this is an expected behavior since every laptop is different and you are running an edited UEFI. In this case, sure, since there are very limited avenues you can pursue to improve thermals, knock yourself out and play around with affinities. It’s not going to help much since it doesn’t actually solve the issue, but you might be able to eke out a bit more performance.

 

[Assimilator]

You bought a laptop.

A device explicitly designed for mobile computing.

A device that makes significant compromises to accomplish this.

If you aren't happy with those compromises, buy a proper PC that doesn't have to make them instead of trying to use software to fix a problem that is not fixable by software.

 

[mql]

…you do realize that you could have just provided a link to that thread from the start? We’re not clairvoyant.
Anyway, so it’s a laptop. This changes/complicates things. From what’s been done and discussed in that thread, it seems the consensus was that both the cooling is not great and that the CPU as configured by manufacturer was running into issues. It’s hard to definitively say whether this is an expected behavior since every laptop is different and you are running an edited UEFI. In this case, sure, since there are very limited avenues you can pursue to improve thermals, knock yourself out and play around with affinities. It’s not going to help much since it doesn’t actually solve the issue, but you might be able to eke out a bit more performance.

You bought a laptop.

A device explicitly designed for mobile computing.

A device that makes significant compromises to accomplish this.

If you aren't happy with those compromises, buy a proper PC that doesn't have to make them instead of trying to use software to fix a problem that is not fixable by software.

Everyone chooses what they want. For this reason, I'm skeptical of criticism that laptops aren't designed for what I do or want to do with them. Or simply put, if the equipment has potential that can be used, then this potential should be available when needed, and not allow some of the available capacity to be lost due to the peculiarities of the software.

I did not limit the Windows processes, but I did limit Dell program processes and a few other minor processes.

This is the current result, the TS Bench test lasts from 10 to 15 minutes.