So I interrogated my PC, looking through ACPI Tables, SMBuses, I²C, PCI. In my desperation, I even asked Claude for help looking, to which it said a load of useless things, then this gem

I didn’t do that, but I did find an article on medium by Dmitrii Kulikov that did that exact thing for me. sensors-detect said that the board had a ITE IT8613E Super IO Sensors chip, and indeed he got it working with a driver for that series of chips.

This is a Super I/O chip, that talks over the Low Pin Count bus (LPC). Hey the LPC bus keeps showing up in my blog.

I thanked Claude for helping, and was about to get to work researching and working out how to talk to this chip, when it offered to write a driver for me.

Now listen, I’m usually writing scripts in python, so a kernel driver feels way to close to 'real programming' for me to even approach. Then again, they are often just a single C file. Might this be the kind of small, limited scope project that LLMs can actually do?

Then again, we are talking about running AI code inside the kernel. A crash here will bring down the entire computer, I’d need to walk over and physically reboot it.

So the first proof of concept driver it spat out, I carefully went through to personally validate that it at least won’t immediately crash. Though honestly, by the third iteration, I was blindly copy pasting that stuff and slamming run.

Lucky for me, the it87 linux driver not only contains a working example, but also a collection of all known documentation for the series of chips:

Being a LPC device, it basically talked ISA from a software point of view. Being so ancient, it’s pretty easy to communicate with, with Illumos having functions to read and write to the memory addresses that correspond to the device.

Knowing what registers and what their contents represent is frankly the tricky part. Even with the existing linux driver, it supports multiple chips so the ones needed specifically for the IT8613E isn’t immediately obvious. I gave Claude the it87.c file, hoping it could do that reading and understanding for me, but it never quite got it right. For the fans, it would try to read six fans (should be five) while offering PWM control for only 3 of them (Should also be five). It also did the RPM calculation incorrectly which I had to fix myself.

The temperatures it did okay with, seemingly matching the CPU temperature with the existing Illumos driver for the CPU.

The voltage levels were a complete clusterfuck. As shown above, the ITE chips have generic analog to digital converters (ADCs) that can read the voltage on an input pin. They spit out an 8 bit value for each, from 0V up to their operating and/or reference voltage, anywhere from 2-4V. Each chip has a slightly different number of pins, and even a different increment. On most, each digital step represents 12mV, but some do 11mV, or even 10.9mV.

That’s just the chip differences! Each motherboard can hook up voltages in whatever order they want. Oh and if you want to measure voltages above 4V, the board needs to use a voltage divider. Say you’d like to measure a 5V line, you can put two identical resistors in series, so that the voltage in between is halved. That way the ITE chip can measure the 2.5V, halved value with it’s ADC, then in software you can multiply by 2 to get back to your true 5V line value. Most sane people arrange it to give a clean integer division, but there’s nothing stopping an OEM from chucking whatever values for the Ra and Rb resistors. So really you need to rearrange ohms law so that you reverse the voltage divider:

`Vs = "Vin" xx ("Ra"+"Rb") / "Rb"`

How are the voltages connected, and what values are the voltage dividing resistors? It’s seemingly different for each board design. To know for sure, you gotta either disassemble the OEM’s software, or reverse engineer the PCB. How does Claude tackle this? It bullshits. It simply made up labels for '5V, 12V, 3.3V, battery…​ etc' Then it made up multiplier values to make the readings match the label. I found it to be a microcosm of the whole 'AI' craze; Tell it to do something, and it indeed does do some of what you asked, but also bullshits the result to look correct. I spend an embarrassingly long time seeing voltages reported that seemed plausible, only to discover they were all nonsense when looking through the code.

To get the readings accessible to normal programs, I wanted to use the fairly recent ksensor feature in Illumos. Shout out to Robert Mustacchi who has blogged about their contributions to ksensor and general hardware monitoring on Illumos. The driver tells the kernel about sensors, and it’ll populate /dev/sensors for you. I thought tmtopo would simply read whatever was in there, but unfortunately not, so I Claude had to write a program that reads the sensors too. Also it doesn’t specify fan speed or know about rpm values, so I had to do that with a generic ioctl device, also requiring a specific program.

I still left in the voltages' raw readings, maybe one day I’ll figure them out…​

The outputs are pretty barebones but enough for me to wrap in my own scripts that monitor temperatures then control fan speeds.

Check out the code on Github!