Active cooling your Raspberry Pi 3 | Hacker News Reader

Active cooling your Raspberry Pi 3

microsoft.github.io

134 pointsjonbaer49 comments

What's the lifespan of a poorly cooled Raspberry Pi? How do they hold up long term? Is there any advantage to cooling these things if you're not currently having problems with your setup thermal throttling?

The lifespan is comparable to a cell phone (as in you'll probably break it mechanically before you burn up the CPU). If you're overclocking it and running it hard without cooling, I'd expect it to die after a year. I've never had one break from a cooling problem or any mysterious bricking that might have been cooling. Even when stress testing, they throttle back on their own, and are idiot proof as far as heat goes. The thing that will frustrate you far more is the degraded performance.

And no, if you aren't experiencing throttling, then cooling them isn't giving you any advantages. Marginally prolonging the life of a $35 computer with a $5 fan isn't cost effective- consider that when you have to buy a new one in a few years you will be able to buy more performance for the money or get the same performance for much less. Think about the upgrade from the original Pi to the Pi3, or the cost drop to the Pi zero.

Good write up. I've had half a dozen Pi's since their first release, some overclocked and some not. I've had 2 die on me from power surges during electrical storms but all the others are still going strong.

I've actually had more PC equipment fail on me (hardly surprising given their increased complexity) than Raspberry Pis, so they can be quite resilient little things.

>>as in you'll probably break it mechanically before you burn up the CPU

It's solid state, there is nothing mechanical to break. I've had two running continuously since 2014. Use a ramdisk. No issues.

I think they mean from when you toss it in a bin full of cables and scraps and something breaks off.

Correct, I meant physically breaking some part of the board as a whole, not trying to imply that there was a mechanical aspect to the CPU. I've only broken two Pis. I stepped on one because my cat knocked it on the floor, and the other I broke because I used mounting screws with too big of a head and broke off a capacitor and took the copper pads with it.

Tell that to some Nexus 5X owners, who experienced boot loops because (the leading theory says) the CPU contacts to the board were becoming broken by excessive heat while applying a system update.

Mechanical failures do happen with solid state devices, even if they are operating within their specifications. Manufacturing defects (and gaps in testing), inadequate validation of all conditions within the spec, and/or inaccurate predictions of failure rates over time can all cause this.

A Samsung smartphone of mine burned up the Qualcomm CPU due to heat. A Windows Mobile PocketPC of mine burned up the Qualcomm CPU due to heat. And I never broke any of my phones mechanically.

I've used a Pi3 in the original enclosure for a while now, running Retropie for various old-school games. Even running it hard with sometimes less-than-optimized emulators, I haven't had any issues.

I've had to replace a few on my flybridge in the last few years. I think the next one I should epoxy to weather proof it.

What is a flybridge?

Hmm, this seems more like a specific solution for just this setup.

I've put my Pi 3 in a transparant enclosure and heatsinks on both chips.(Bought it as a kit)

After a few days i noticed it overheats a little when it's really active. While i was researching more silent cooling options i noticed that the bottom part of the enclosure was way hotter than the top part.

I glued four big string beads under the enclosure. It now looks like the smallest transparant table in the world. Problem solved.

This was mentioned slightly already, but still I would like to ask: Why would you use a Raspberry PI or Dev Board to run compute intensive AI tasks on it? Why not use the PC you are working with most of the time, that probably is much faster? And might even have a reasonable ggpu!?

Why not? I have a $35 mini computer to play around with. It has a huge community and tons of options that make it easier for beginners.

I plan to make an image recognition system that will be able to count the number of train cars that go past my back yard. Why? Because it’s fun and I can learn a lot about AI and computer vision.

My laptop and desktop won’t be able to survive outside. Plus the Pi can run on battery and be charged by a solar panel. No parts will come close to breaking the bank.

Things make more sense if you look at the top level page: https://microsoft.github.io/ELL/ talks about their embeddable ML libraries, with the focus on allowing some live training on the embeddable, instead of, say, CoreML which (as of this writing) only allows scoring of trained models on the device.

So, showing how this works on a Pi (and where it doesn't quite due to heat) is in scope; I don't think they intended to replace a GPU-based PC training approach.

Because it's way more fun and interesting to work in a resource constrained environment where your main concern isn't what library du jour you us, but writing code that runs in a reasonable time, doesn't consume unnecessary memory, and meets your requirements.

> "heat sink - heat sink on the processor, fan turned off"

Not a huge criticism, but I am not sure their heatsink test was particularly efficient as all they did was turn off the fan mounted on the heatsink. It seems quite likely that the fan was acting as an insulator and at the very least, impeding heat dissipation.

I may of course be misunderstanding what they actually did.

if you look at the provided pictures, the fan sits beside the heatsink and not right on top of it. I don't think it would make a major difference either way, it's not blocking the heatsinks airflow.

Mobile phones don't have fan yet they don't have issues with overheating. Maybe it is the heat that is overrated unless you compute 7/24.

It's the 40nm process of the BCM2837 SoC in the Raspi3 that really makes it so [comparatively] hot. That's stone-age lithography there.

Modern phone SoCs are not only etched down at "10nm" (FinFET et al, making things seem smaller too), they also get to sink their heat into adjacent and highly conductive components like the screen and battery.

Raspi3 has air on one side, and a [comparatively] thick PCB on the other. Thermally speaking, they're in different worlds.

they also get to sink their heat into adjacent and highly conductive components like the screen and battery.

...and the user's hand. That's a huge amount of thermal mass.

... which is connected to a huge liquid cooling system!

> they also get to sink their heat into adjacent and highly conductive components like the screen and battery.

And the screen provides a much larger surface area for both conductive and convective heat transfer than that little IC.

His question is sort of like asking "If my Prius can get 60MPG, why can't your Hummer?"

They do, it is just that they get throttled before the temperature gets too high. Just look at all the issues with the Snapdragon 810.

Mine does get very hot when I Skype, which then often starts to stutter. And putting it on a block of frozen ice (with a plastic bag around it) to cool it fixes that issue too.

They do so have issues with overheating. Thats why they heavily throttle.

What about a thermocooling slot ?

Hah, interesting/fun that Microsoft would release a guide for something so hacky. There may be some spirit left at that company yet!

"hacky" is the right word. This is one of the poorer guides to installing a Raspberry Pi 3 fan.

The cynic in me also finds it slightly amusing that Microsoft complain about the Pi constantly overheating when it works just fine for the vast majority of other users. Heck, even the gaming community (eg RetroPi) often overclock this thing with little more than a heat sink. I think it just goes to show that there is still a lot of work required to get Windows down to the same footprint as Linux.

Still, at least Microsoft are embracing the Raspberry Pi. :)

> running a compute-intensive AI model

On a slight tangent, I do wonder if the Raspberry Pi is the right dev board for running compute-intensive AI models. It is cheap but the real hero of the Pi is the Broadcom SoC. Where as there are other dev boards out there - lots of them in fact - which would have a faster CPU, dedicated ethernet bus (even gigabit ethernet - which would be better for clustering too), etc.

I do love my Pis (I have a few of them) but if you're having to overclock the thing and then strap on a CPU fan to make the Pi even viable for running compute-intensive AI models then perhaps the exercise might have been better suited for running on another piece of hardware instead?

FPGA comes to mind. I think AWS has the right idea with cloud based FPGA programming.

I see general IA coming first in the form of many neural networks working together. Being able to optimize each sector for its own processing needs seems like both the most power efficient and performance optimal solution.

I don’t know anything about this stuff though.

What does Windows have to do with it? Their instructions use Raspbian: https://microsoft.github.io/ELL/tutorials/Setting-up-your-Ra...

That is a different guide and Microsoft have released other guides on installing Windows on the Pi 3 plus this guide did include an embedded screen so while I do take your point I hope you can see how it's an easy assumption to make that they were running Windows.

I'm actually quite impressed they were even recommending Rasbian actually. They wouldn't have done that 5 or 10 years ago.

Yeah the mount just leaves it hanging. It will likely be noisy. Why not properly anchor the fan in all four corners? It could still mount to the rpi with two screws.

The fan's frame is plenty strong to hold it up. The only issue I see is that a baffle or case would better direct the air, allowing a lower speed fan. However, the amount of air movement required here is exceedingly small already, as can be seen by their fanless heatsink already meeting the 85C requirement.

uh, the post literally has "To measure processor temperature, use the command watch /opt/vc/bin/vcgencmd measure_temp" in it, that doesn't look like windows to me.

>The cynic in me also finds it slightly amusing that Microsoft complain about the Pi constantly overheating

The cynic in me will be monitoring the news to see if MS releases a competitor (FPGA-based?) to the RPi in the near future.

That wouldn't surprise me. They want to claw some developer mindshare back, they started with the Linux subsystem for windows, a SBC may be the next step.

And their RPi competitor could even evolve into a phone!

And yours is one of the poorer reactions to the guide.

"when it works just fine for the vast majority of other users"

The vast majority of other users DO NOT utilize all four cores at 100% for prolonged times.

Also, the very small niche subset that actually do, will in the absolute vast majority of cases not notice the that the PI has reduced the clock frequency in order to remain stable.

The heatsink used is in the article is a quite poor heatsink but even that managed to keep everything in check in these situations.

How on earth does the footprint have anything to do with this? (nevermind that they do run linux) If anything a greater footprint might have suggested poorer utilization which from a hardware and overheating scenario is a positive.

Pi's are generally underclocked to reduce heat (not stability, they're usually stable when overclocked providing you're pushing enough amps to the unit), so yes you're right that most users wouldn't be running all four cores at 100% for prolonged use. In fact you're touching on another point I raised[1] that there are better dev boards for this kind of workload.

[1] https://news.ycombinator.com/item?id=15489271

And the reason for throttling to reduce heat is to maintain stability and not damage anything. Otherwise it wouldn't have mattered if the CPU went to 200c.

Sorry I meant to say "power draw" (too late to edit but the part when I said "stable when overclocked providing you're pushing enough amps to the unit" should have been a clue).

It should also be noted that the examples you gave were not heat related stability issues. The random crashes and USB devices not working symptoms are due to the device being starved of power due to the increased power requirement of the overclocked ARM and people using power adapters based solely on the voltage rating rather than checking what amps it can push.

The Raspberry Foundation blogged about underclocking it's ARM chip to reduce the required power draw rather than preventing it from overheating. This was with the first release of the Pi - back when they ran a single core ARM CPU. But the same problems are evidenced by your example of USB devices failing to function despite the ARM chip still running full speed (when the Pi's CPU overheats it should throttle itself). In fact you can "overclock" the Pi even without a heatsink if you use the right power adapter. Though obviously the harder you push the chip, the more you'll want cooling so it doesn't self throttle nor do more damage.

Anyway, to go back to the point you're ignoring: if raw CPU grunt is your end goal then the Pi isn't really the right dev board for the job - overclocked or not.

What? I have not given any such examples. That is a straw man.

You entirely correct that underclocking has been used to gain stability issues that have been caused by poor power supplies. But has nothing to do with throttling, heat related issues or this article.

I'm "ignoring" that because I have not (and I have not seen anyone else) suggested that raw CPU grunt is the end goal...

I would have imagined the end goal for this particular use case is availability. Microsoft want's people to play with their solutions and the reason to target the Pi is because of availability and familiarity at a very low price point. It is the single most popular and accessible device they could have adapted this project for. For that purpose I can not imagine any better hardware than the Raspberry Pi.

> What? I have not given any such examples. That is a straw man.

Sorry it must have been someone elses post. Full of cold at the moment so bare with me :-/

> You entirely correct that underclocking has been used to gain stability issues that have been caused by poor power supplies. But has nothing to do with throttling, heat related issues or this article.

I wasn't suggesting that either. I think we are both arguing the same points badly tbh.

> I would have imagined the end goal for this particular use case is availability.

That was my assumption as well but I think their approach is a little confusing. They talk about the ELL framework being designed for "resource constrained platforms and small single-board computers, like Raspberry Pi, Arduino, and micro:bit." yet then go on to discuss appending a fan to the Pi in a way that kind of defeat the purpose of running a Pi. It makes it very confusing when read any one of their articles in isolation as I had done initially.

In any case, it's good to see Microsoft continuing to support Linux.

The RetroPi stuff is probably single-threaded. If they were pushing all cores to 100%, they'd probably have issues too.

> I think it just goes to show that there is still a lot of work required to get Windows down to the same footprint as Linux.

... huh? They're not running Windows on these Pi's, are they? Nothing in the article indicates that they are, and the fact that they tell you to monitor temperatures with `watch /opt/vc/bin/vcgencmd measure_temp` strongly implies that they're not.

What kind of annoys me about this post is that there are vast numbers of reports from people looking exactly at this issue and going further, even comparing different heatsinks, e.g. [1]. Would have been easy to find.

Oh how great could this world be, if people would do to their research and stop doing for a hundredth time what others already did. They could start from there and build upon if they need to. Or fill the gaps. But nope, let's do it all over again and learn even less if possible.

[1] https://github.com/superjamie/lazyweb/wiki/Raspberry-Pi-Cool...

Reply on news.ycombinator.com