Show HN: My GPU Fan Saga – A DIY ATX Fan Controller | Hacker News Reader

Show HN: My GPU Fan Saga – A DIY ATX Fan Controller

shafq.at

23 pointsashafq15 comments

Could have been simpler still. You can read the internal temperature of the ATiny85 [1]. You can also get the ATiny85 to speak USB and even be flashed via USB via the micronucleus bootloader [2].

[1] https://andrey.mikhalchuk.com/2011/06/20/reading-attiny85452...

[2] https://github.com/micronucleus/micronucleus

I did try the internal temperature sensor, but it didn't measure the ambient temperature. I have read that if you set the ATtiny85 to sleep mode for some time, it will converge to ambient temperature, but it would take too long to validate, and I didn't want to go down another rabbit hole.

USB Bootloader is interesting, but I already had a few ISP loaders. You can make an ISP loader with Arduino as well. [1]

[1] https://docs.arduino.cc/built-in-examples/arduino-isp/Arduin...

> I did try the internal temperature sensor, but it didn't measure the ambient temperature. I have read that if you set the ATtiny85 to sleep mode for some time, it will converge to ambient temperature, but it would take too long to validate, and I didn't want to go down another rabbit hole.

It's kind of better than ambient, it would reach the temperature of the GPU if thermally coupled. The sleep-mode thing would just be a case of lots of micro sleeps to discount the temperature of the chip, could have even under-clocked too.

> USB Bootloader is interesting, but I already had a few ISP loaders. You can make an ISP loader with Arduino as well.

The benefit of USB is being able to get debug data from the fan controller as it runs, and be able to re-flash it whilst in the PC by attached to an internal USB bus.

This way you could also log out what the GPU reports, vs what you measure, vs what the ATtiny85 measures.

How does the internal temperature sensor help here?

I think he means tape/glue it to the heat source to (mis)-use it as a probe. "It's not stupid if it works" but I think there would need to be a few deg of compensation to account for the thermal resistance of the package.

Doesn't even need to be attached to the heat source, just thermally coupled. Could be achieved for example via a short, fat, insulated, shared ground plane or something.

The compensation might be needed, and the internal thermal sensor also needs tuning in any case if you read the specs, but it's close enough that you could use it out of the box and only be out by a few degrees. Chip temperature with regular sleeps isn't going to be a significant factor.

There is enough code space on the ATtiny85 that you could actually build a model that takes measured temperature T at time t (T_t), the goal temperature of G (G_t) and produce a predictive model for M(T_t, G_t) = T_{t+n}. You could then do something a little smarter than a PID controller that takes in some profile considerations, i.e. slow (low noise, low RPM, likely for normal usage), fast (high noise, high RPM in bursts, likely for gaming).

Then you have the problem of the PCB getting in the way.

The author has an I2C temp sensor next to a giant MOSFET. The author also uses clearly ChatGPT generated avatar and mentions I2C is bit banged yet completely left out circuit schematics which incidentally includes freewheel diode for the fan. PCB getting in the way isn't going to matter in this instance.

> The author has an I2C temp sensor next to a giant MOSFET.

Yeah... It's either overkill, or there is some decent power going through it and it will affect thermal measurements anyway. Not to mention what it might do to the reference voltage.

> [..] mentions I2C is bit banged [..]

They possibly bit banged it? But it wouldn't have been needed given the hardware support...

> [..] yet completely left out circuit schematics [..]

Yeah, which is far more important that anything else.

> [..] which incidentally includes freewheel diode for the fan.

This may have been sensible to be fair, especially if you happen to rapidly change the delta PWM signal, you could end up with a decently sized back-EMF. But I think even better would be to sink it into something, rather than back into the power supply of sensitive components...

> PCB getting in the way isn't going to matter in this instance.

Agreed.

I have been wanting a PWM controller for my Tesla cards in my workstation. Currently I open the case and manually turn up their fan speed when I’m going to use them, and turn it back down afterward. What I would like, but am not very interested in spending the time on, is a microcontroller that controls the fan speed based on messages it gets over serial from a script running on the host that tells it what speed to run based on GPU temperature. When I first set it up, I also placed thermistors that the microcontroller could use as a fallback if it doesn’t receive commands.

Seems like something similar should already exist, but I haven’t been able to find anything that is a close enough fit.

I have a pretty custom GPU cooling setup on a few machines (I run ML workloads locally and I want stuff to be quiet).

Couple of gotchas that I ran across. I found that on Linux, desktop PC fan control support is pretty abysmal. The sensor library that everyone relies on, lm_sensors, is semi-abandoned and didn't recognize sensors on my relatively popular, 7 year old ATX motherboard and GPU. It also requires having Perl installed.

About GPU cooling in particular - modern NVidia cards in particular seem to have a built-in minimum of 30% fan speed when controlling them manually. The connectors are also a different, smaller connector (perhaps a JST PH?).

> The sensor library that everyone relies on, lm_sensors, is semi-abandoned and didn't recognize sensors on my relatively popular, 7 year old ATX motherboard

This is mainly because the companies that make the “super io” controller chips (ITE and others) are deluded and think that they have some sort of cutting edge tech and their chip documentation must be super secret.

Years ago there was someone at a mobo manufacturer who would leak the datasheets, but they left or got found out, and there hasn’t really been a source for that info since.

The Windows FanControl software uses LibreHardwareMonitor as the sensor backend. It works pretty well in my experience.

Would be nice if you posted a picture of the schematic in addition to the PCB rendering.

Thank you! I will add them later on. You can access the schematic and PCBs here:

GitHub: https://github.com/Colahall/tiny85fancontrol/tree/main/schem... Codeberg: https://codeberg.org/Colahall/tiny85fancontrol/src/branch/ma...

Reply on news.ycombinator.com