Linux Developers Discussing Possible Kernel Driver For Intel CPU Undervolting

[Kayden]

Which Intel developer that come here we can ping about it? Kayden

I have no idea about this stuff or who to talk to, sorry! :(

[tildearrow]

It's not that the AMD CPPC driver isn't 'allowed', it was that issues were pointed out and ultimately AMD decided to abandon that effort for now to push their resources elsewhere.

There is a typo... unervolting

[geearf]

It's not that the AMD CPPC driver isn't 'allowed', it was that issues were pointed out and ultimately AMD decided to abandon that effort for now to push their resources elsewhere.

Didn't the guy leave AMD anyway?

[hiryu]

I'm using "intel-undervolt" on my laptop and it's working just fine. I'm not on that machine right now so I can't confirm that's the name. I think I have a stable -90 mv undervolt working in Linux on my Lenovo Legion 7i. Unfortunately, audio output from the built-in speakers still doesn't work, but that's a separate topic.

If there's a more... "official" or "endorsed" way via the kernel, I'd be happy with that too I suppose.

[moriel5]

Which Intel developer that come here we can ping about it? Kayden

I have a friend (and former co-worker) who had joined Intel's CPU engineering team in my country a few years ago.
I have no idea whether he reads on Phoronix, though I do not believe he has an account on the forums, even if he does read here.

I'll try asking him about this.

[Sonadow]

What is foolish about wanting less power use, heat, and noise? .

It's foolish because it can already be achieved on Linux by manipulating stuff inside /sys/devices/system/cpu/intel_pstate/<relevant files> without messing around with voltages.

[ed31337]

I have a lot of experience with dynamic undervolting in microcontroller electronics. It is a very lucrative endevour for battery powered devices, but it does require a bit more engineering time to make sure it all works safely without data corruption. In my case, I was able to improve my application running for a few hours on a single NiMH AA to being able to run for several days!

Most of my power savings revolved around undervolting an SD card. My app required bursts of data from the card on a regular basis, but in between bursts of data, there was silence. Originally I thought I would just completely power down the SD card between bursts of data, but that turned out to be a no-go because the SD card requires quite a long boot up time before it can be accessed again.

However, during periods of no activity, you can often safely undervolt a device and it will retain it's current existing state okay. When you need to communicate with it and make it do active work, you need to bring the voltage back up before you start doing anything, else you'll lose state and corruption results (not to mention the difficulties in dealing with wrong I/O voltage thresholds if you're undervolted). In my case, waiting for the voltage to rise was a lot less time and electricity wasted than waiting for the SD card to boot back up from zero, so undervolting really helped a lot for me.

You can get some interesting compounding benefits from reducing power consumption. Obviously, reduced power means less heat produced, so if you get low enough you end up not needing to power on cooling fans, which saves more power. But in my case, there was another interesting effect in that once I got down lower and lower in power consumption, the internal resistance of my battery became less and less and the efficiency of my switching power supply got better and better, which really helped extend my run times even further. It was great! I remember at one point making a small optimization to my code and finding a much greater improvement in run time than I was expecting, due to the compounding benefits in efficiencies elsewhere.

[DanL]

It's foolish because it can already be achieved on Linux by manipulating stuff inside /sys/devices/system/cpu/intel_pstate/<relevant files> without messing around with voltages.

Generally, undervolting saves more power than underclocking.

[woife]

I have a lot of experience with dynamic undervolting in microcontroller electronics. It is a very lucrative endevour for battery powered devices, [...].

Thanks for sharing. It was an interesting read and I learned a few things

[Alex/AT]

Just providing something like msr.allow_userspace kernel option would do for tweak happies, who are ready to go with the risks - they are outspeccing the h/w anyways.
For the mainstream, no, this is definitely not viable.