coil whine

[djacobow]

Friends of mine who run GPUs for gaming and machine learning have told me about "coil whine". This is audible noise coming from (presumably) inductors in the DC/DC converts on the graphics boards. Seems to happen more when the machine bogs down and current use increases.

I guess this surprised me because I figured that such converters run well above audio frequency. Either they don't, or there is some kind of sub-harmonic thing going on. I can imagine electrical sources such as a control loop that creates oscillations much lower than the PWM frequency. Can also imagine that the actual current use of the card is oscillating at audio frequencies when the code loop is just so. (epicycles in the code cause regular audio-freq load variations)

Are there obvious mechanical ways to get a lower frequency than you put in?

Curious what a user can do about this (probably little) and what a designer would do about it?

[jeremy]

Sub harmonics are common in older chopper stepper motor drivers with microstepping (steppers are just big coils after all). For example, DRV8825 which has a fixed but non-audible frequency but almost always (imo) has audible noise at idle. Solution is to use a system with a dynamic frequency that can be adjusted away from a resonant point. Or just make the frequency so high that sub harmonics are still inaudible.

[SiliconWizard]

Also, not all DC/DC converters run at fixed frequencies. So you would get some kind of modulation based on the current draw profile, some frequency components of which can lie in the audio range.

One frequent misconception is to think that the noise actually comes from the coils (inductors). This can happen for sure (in which case, potting can be an option, but beware of heating issues), but is not the most common source of parasitic audible noise in running DC/DC converters in my experience. Much more common are ceramic capacitors that vibrate at high frequencies (piezoelectric effect).

Some pointers:
https://www.edn.com/design/components-and-packaging/4364020/Reduce-acoustic-noise-from-capacitors
http://www.samsungsem.com/global/product/passive-component/mlcc/acoustic-noise/index.jsp
https://product.tdk.com/en/contact/faq/31_singing_capacitors_piezoelectric_effect.pdf
https://www.murata.com/products/capacitor/mlcc/solution/Naki
http://www.eenewsanalog.com/content/reducing-mlccs-piezoelectric-effects-and-audible-noise

[PA4TIM]

I have a coil saturation tester, when I increase the current to much the coil saturates and start whining. http://www.pa4tim.nl/?p=1859 for more info

The core of an inductor is a self inductance "multiplier" the amount is stated as AL.
If the current through an inductor with some core gets to high the core looses his function (the magnetic field becomes to high), the AL drops and the selfinduction too. This results in more current that cause even less self-induction etc until something dies.

An thing that can happen is movement of parts producing soundwaves. I had a Murphy receiver that had an audio transformer who thought he was a loudspeaker. And rather loud  too.

And smps frequencies can drop into the audio frequencies but we can not hear audio frequencies in the form of EM fields, something must turn them in to audio waves first  ("air pressure" variations, like a speaker does)  That can be an transformer/inductor and maybe some types of capacitors can do this too. But as far as I know they drop so low only around idle, not under full load.

If the computer works fine then I think something is resonating/vibrating and so produces audio. If it saturates the videocard runs out of juice and I think that will be visible or noticeable. If you can reach it you can find the part using a piece of tubing as a sort of stethoscope. I made one using an Y splitter and the rubbers from those in-ear-headphones. I made that a while back to find the cause of a ticking sound in a calibrator, turned out to be a cap that was breaking down under certain conditions.
If you have the space you can also use a screwdriver. The handpiece to your ear, the tip on the component. I used that while working on motorbikes and cars to listen to bearings.
 

[Distelzombie]

My laptop does the same thing. But in that case it is somehow connected to audio output. But the whining doesn't come over the speaker, instead from somewhere on the board.
Sometimes when a audio clip ends, the whining noise is extremely loud and only goes away if I start another audio clip. Not the slightest idea why that happens or what it really is. But it's still there when I disconnect all speaker.

[ejeffrey]

Are there obvious mechanical ways to get a lower frequency than you put in?

Basically any non-linear element can do AM envelope detection.  This can be from rectifier diodes, non-linear piezoelectric coefficients, magnetostriction, or I x B force in inductor wires (since B is proportional to I, the force on a wire is proportional to I^2).  So if there is any supply or load variation at audio frequencies those non-linearities can rectify the signal from the high frequency PWM and generate audible tones.

If you hear the noise primarily at high load, these are the issues I would look for.  Some circuits make noise primariliy at low load, in that case it is more likely a regulator going into pulse skipping mode so the frequency drops.

Lots of things in computers happen at audio band frequencies.  Timer / scheduling interrupts, video frame rates, network packet reception, USB bus polling, and many others are typically at 100s of Hz to kHz and have harmonics all through the audio band.  All of these can create audio frequency ripple in the load current.

[CopperCone]

Do they have metrics for winding coils in geometries that focus on accoustic resonance dampening rather then electronic parameters?

Like an anti speaker.

[Miyuki]

If you hear the noise primarily at high load, these are the issues I would look for.  Some circuits make noise primariliy at low load, in that case it is more likely a regulator going into pulse skipping mode so the frequency drops.

Lots of things in computers happen at audio band frequencies.  Timer / scheduling interrupts, video frame rates, network packet reception, USB bus polling, and many others are typically at 100s of Hz to kHz and have harmonics all through the audio band.  All of these can create audio frequency ripple in the load current.

Most of it is power management caused
Modern chips work in short high frequency pulses and then stop for short time to cool down
And also that interrupt think for example some computers have audible noise when at idle move mouse as they wake cpu from no load and almost no power consumption and go to short high speed/power pulse and then back to idle

Question is can you really have totaly quiet power supply when taking 200A pulses at audible rate, I think at this huge currents you make forces even at bare pcb traces

[chris_leyson]

@Blueskull You nailed it in one - pulse skipping. I'm not sure if it applies to GPUs it all depends on the converter control mechanism. Some low power offline flyback controllers will go into this mode of operation at low power levels. Recently been looking at Power Integrations LNK626/625 for a low power flyback, it's uses a bang bang hysteretic control loop with more or less fixed peak primary current and uses pulse skipping at light loads. As you increase the load the spacing between groups of pulses gets shorter and the frequency of the pulse grouping works its way through the audio band.
If you've got a flyback transformer that hasn't been varnish impregnated too well you can hear the core or sometimes the windings themselves. Large X caps will also generate some audible noise believe it or not.

Only the other day I built up a pcb with an LMR23610 buck converter to drop 24V to 5V, it's not driving much just an ATmega16U2 and an RS485 driver. I had a few spare LMR23610s and I was interested in looking at the low load efficieny so that's why it got designed in. Anyways, I hooked up the board and ran some data back and forth at 9600 Baud and heard the faint sound of a modem. It turned out to be pulse skipping because the load current is so low and I hit the sweet spot. The buck inductor is a Wurth WE-PD-1050 and it's working really well as a small loudspeaker, I didn't expect that. Will have to record it and see if I can extract any valid data. 

[Cyberdragon]

I have an old CRT based machine that runs the flyback at 3KHz for some reason. The manual states to listen for the "hum" to see if it's working.

The sound of a tarantula landing square on Justin Bieber's face must be a 3KHz "hum" to them. (Granted, I have the case off, but still unpleasant)