292 comments

[ 2.9 ms ] story [ 280 ms ] thread
I'd guess that heatsink and a case with heatpipes would be sufficient for most, though perhaps not all, uses and silent, which for me is pretty important on a Raspberry Pi
I tried a heatsink but it didn't really help unless I had a desk fan providing some air flow.
This pretty much lines up with my findings. I popped the pre-installed heat-spreader off and applied a heatsink with noctua thermal paste... basically no difference- step-by-step chaos here: https://twitter.com/Gadgetoid/status/1151442153584910337

Just adding a small fan (trying to avoid blowing our own trumpet too much here) absolutely demolishes the cooling performance of any reasonably-sized heatsink. Presumably there's just not that much passive airflow over something an inch square, even with convection.

I still wonder what would happen if a full desktop-sized heatsink was coupled to the Pi 4's SOC.

   I still wonder what would happen if a full desktop-sized heatsink was coupled to the Pi 4's SOC.
In all likelihood, you'd see diminishing returns as you get bigger, but that would be quite the sight to see.

I've been tempted for a while to build a big case with some sort of mounting standard (maybe a Peek array?) so that I can add whatever hardware I want to it (mostly electronics). This seems like a job for that.

convection cooling needs a decent gap between the fins for it to really work. This is why something like Noctua's passive cooler ( https://www.anandtech.com/show/14486/noctua-shows-off-concep... ) looks totally different from their tower coolers with a fan.

Just slapping a normal CPU heatsink on the pi4 will probably work OK just do the thermal mass and exterior surface area of it, but you're probably not going to get much convection-cooling from it, either.

I appreciate you actually trying. A "negative" result is still valuable.

What I don't like about the article is that small fan have to spin fast (= noise) to generate enough flow. I'd much rather use a 80 mm fan, slowed way down, but I suppose I'd had to roll my own case for that. Better still, wait for someone to start offering this (or maybe this: http://blog.flirc.tv/index.php/2019/06/24/new-pi-4-cases/)

I'm not the best judge, but these are really quiet fans. I can hear the ones in my laptop, but don't notice Fan SHIM at all. We've had a couple of customers who have, but a replacement fan has been quiet- guessing it's a tolerance issue with the plastics.
did you try a big enough heatsink? I put a northbridge heatsink on my odroid xu4, and it did even better than the stock cooler with a small fan.
The problem is there are now a number of chips on the board—many of which are too small to pop a heatsink on them—which get crazy hot. If you have the Pi inside any kind of enclosed case, that heat just bakes everything else, even if you have heat sinks.
You could design a copper heat-spreader to pop over the heat-sensitive part(s) of the board, and then apply a heatsink to that. Maybe copper foil would do the job.

Another approach is to just run the Pi submersed in mineral oil, with liquid convection then taking the heat away from the board and shedding it to the surrounding environment. We usually don't run our electronics like this (even though the general approach is widely used for cooling needs of all sorts) because mineral oil is gross and might even impact the endurance of our hardware - but the Pi is tiny and cheap enough to make this a non-issue.

The Pi has connectors on both the long and short edges, so you'll want to be sure everything is plugged in before you dunk it.
That's what Apple used to do with their Macbooks. Heat pipes into the metal frame which conducts it over a larger surface area creating natural convections which carry the heat into the air.

The real downside is that mass plays a substantial role. Meaning thicker metal, larger cases, are better than thinner/smaller. Whereas a fan can accomplish similar results for less cost and space (but not passively).

If people don't think passive can work they need to check out Apple's recent Mac Pro that passively cools two Radeon Pro 580Xs. There's no limit, except cost.

Personally I still feel a combination of heat pipe, into heat sink, with a small progressive fan is the ultimate. If it is large enough the fan should stay off except under heavy load.

While there are not extra fans on the cards themselves, I'd hardly call anything in the Mac Pro "passively cooled" considering there are three giant fans on the chassis immediately in front of those heatsinks.
Apple has a pretty long history of passively cooled devices, including the G4 cube, which had a massive (for the time) CPU and GPU, if you wanted in it, was 8 inches cubed in volume, and cooled convectively.
I can recommend Pimoroni's fan shim. They provide a python project that allows you to set temperature thresholds to kick the fan in so it's not on all the time. It's not silent but it's pretty quiet.

https://github.com/pimoroni/fanshim-python

I thought about getting one of these, as they are a little more turnkey... however I couldn’t find an easy way to purchase and get quick shipping in the US. Is there any other way besides ordering through the UK?
Ahoy! We ship (pretty fast!) to the US. :-)

Failing that I'd hope Adafruit will have them in stock soon! I'm one of the co-founders of Pimoroni, drop me a line if you need any other help.

(comment deleted)
Indeed, the fan shim is insanely cool and I just bought one of these the moment that it was back in stock together with my Raspberry Pi 4. It also works on the Pi 3 and 2.

Here's the link if anyone's interested: https://shop.pimoroni.com/products/fan-shim

Glad you like it! We think it's the best solution for cooling any model of Raspberry Pi that's available.
Does it have a standalone mode, where it doesn't need to be controlled by software? I'm interested in using it with prebuilt images that wouldn't be able to run the python to control the fan.
By default it's always on.

If you (optionally) install the daemon then it can use the Pi SoC's internal temperature sensor to control dynamically based on need.

I don't suppose it has a fan RPM sensor or anything, so one can monitor the health of the fan itself?
@mindcrime - I can't reply to you due to max depth. The current version doesn't, we're having a new, custom, fan made with an RPM feedback line for a future model but it'll be a couple of months yet.
If you click on the time above the post, it'll take you to a page where you can type a reply.
... we're having a new, custom, fan made with an RPM feedback line for a future model but it'll be a couple of months yet.

That's good to know. Thanks for the update!

(comment deleted)
Insanely cool! Very convenient form factor too. You're about to sell two more of em.
Never seen these before and just bought one! I'm going to try running some old 3d games via Wine on the Pi and this fan will probably really help performance. :)
Awesome. I didn't know about these until day, and I just ordered one. I may need a couple more down the line as well.
Cool fan but it says "not heatsink compatible", so if you are like me and attached the heatsinks that came free with your Pi case, you will be out of luck.
I second this. The fan is amazing!
So does the pi 3 if you stick it in a case, even with a heatsink.

I was using a pi 3 with a heatsink in the official case to play a h265 movie (which will be software decoded on a pi3). After about 10 minutes I noticed it started dropping frames and it displayed the thermometer symbol in the top-right of the display.

When I removed the top lid from the case, the temperature dropped enough for the thermometer icon to disappear and playback to continue smoothly.

Yeah, thermal insulation behaves in much the same way as electrical insulation. If you put a conductor in series with a resistor, the conductor doesn't conduct all that well.
Seems like the winning solution would be to do as Apple does and make the case the heatsink. Aluminum is a pretty good thermal conductor, and at the very least it'd get you lots of surface area.
Wouldn't that make it hard to touch?
It wouldn't be too bad, the heat would be spread out and the surface would be at a much lower temperature than whats actually measured on die.
Over the years I've used Raspberry Pis as little media players, but they kept burning out every 3-4 months; the card reader (their SSD) wouldn't work, or they'd just blink some other error message. I thought they were just cheaply made ($35 after all). Anyway, long story short, I got a fan and the last one has been working for over a year without issues.
SD cards have a limited number of writes so if you have the OS constantly writing to it then eventually it sets itself read-only.

You can mitigate this somewhat with https://github.com/azlux/log2ram

With some finagling you can get them to boot an OS from an external disk then your SD card can just hold a read-only boot loader.

PLUS: You can get MicroSD card file system corruption if not using industrial grade MicroSD cards - especially when losing the power...
I believe this is mainly down to the FAT16/ FAT32 filesystem on the SD cards. Those filesystems have never been particularly reliable, when those were the dominant hard drive formats, MANY users back then lost data (and backing up to floppy had its own set of issues).

This is why you have to unmount USB drives, if you pull the drive while data is still in cache would also create problems. The difference is when copying files to the drives, it's sequential, you may lose a file or two. When it's the OS using it as the main drive there are loads of write operations going on.

At least it's easier to add battery backup to a Pi then most other computers.

The microSD card in a Raspberry Pi uses an ext4 filesystem, not FAT.
The root partition is ext4, while the boot partition is FAT; plug it into a windows machine and it just sees the boot part, so people assume the whole thing is FAT
> I believe this is mainly down to the FAT16/ FAT32 filesystem on the SD cards.

Well I lost some to power failure, and they all had ext4 on it. Anecdotal I know, but I'll be avoiding RPi for anything that needs to persist data in the future.

In the past when I was using a rasppi (v1) as my media player, I used the SD card only for booting off of. The media files themselves were persisted on my nas, served over nfs. Using network storage really seems like the way to go, particularly now that the shared bus issue has been resolved. At the very least, real hard drives are much more efficient (byte/dollar) than SD cards.
Also if you are not using a good power supply the Pi will happily operate on less current except when it needs that current during power-intensive operations like watching a movie. This could be causing silent corruption as it happened in my case. If you ever see a thunderbolt symbol best check your power supply.
Flirc makes excellent passive cooling cases for RPis. Basically the entire shell is a heat sink.

Works for me, all my little media boxes run happily with them, despite running 24/7.

Same here. I have 2 Flirc cases for the RPi 3b, and it keeps it happy. One is used for media the other for emulator.
If anyone is interested in putting in a bulk order for Flirc Pi4 cases to save on shipping, let me know. I am in the San Diego area in SoCal.
Well the pi3 does not support h265 acceleration natively, it should have been better to use it on h264 movies.
Hardware acceleration comes in degrees... It's totally possible to hand-code some GPU firmware to decode any bitstream far more efficiently than the CPU can do it...

It's just a lot of work, and either Broadcom hasn't done it, or they don't want to license it to the Raspberry Pi.

I thought (for some reason, I could be wrong), that hardware acceleration for h265 implies actual dedicated silicon, either in the SoC itself or on the board?
That is correct. A pi 4 can doe h265 acceleration, but for anything other than a pi 4, use h264 content.
I'm hosting a Node.js webserver from my RPi3, I have heatsinks and a case installed with no fan.

Seems to run fine, it's never overheated or throttled for me.

I've a Pi3 with a CPU heatsink, in case. I have not seen it throttle after adding the radiator.
I installed libreelec on my Raspberry Pi 4 the first day I had it. The CPU immediately started (while playing 1080p video) to heat a lot to the point it stalled. Since I haded a motherboard chipset heatsink that I had laying around on the CPU it worked flowlessly. (I keep it without any case)
This was exactly my intended use case - as a gift for relatives - the problem I've had is that I need it enclosed to prevent tiny little fingers from getting hurt/destroying the thing. Tried soak testing it one night and experienced the same thing :/

Currently I'm hoping that I can modify the case this weekend to fulfill my needs.

As something intended to be used by kids to learn about computing I do wonder how many are going to get hurt fingers from the heat output by accidentally maxing the cpu and/or touching the USB ports, etc.

> As something intended to be used by kids to learn about computing I do wonder how many are going to get hurt fingers from the heat output by accidentally maxing the cpu and/or touching the USB ports, etc.

Sounds like one of the unexpected lessons will be heat dissipation :)

Probably outside of the realm of weekend hacking, but I wonder if something like this would work for you: https://flirc.tv/more/raspberry-pi-4-case

Best of luck!

Hadn't come across the case. I'll give it a go if my DIY attempts dont work (or are crap) :D
Even the previous-gen Pis would easily throttle under load without a heat sink (although usually installing a decent heat sink was enough to resolve the problem, assuming decent airflow). As the article notes, the heat was mainly from the SoC in the past. The Pi 4 is on a whole different level indeed.

I imagine the logic here is the same as with many other Pi accessories: if you need it, you'll buy it or get it as part of a bundle. In some cases throttling is not a huge issue. But we're a far cry from the simple plug-and-play Pis of the past... Another issue is power -- you can't simply run the newer Pis reliably off any old cell phone charger you have laying around.

That’s just it, though. The official Pi power supply is adequate and good for powering the Pi in any condition, but the case is a far cry from adequate. At best, it just toasts everything inside more evenly, at worst it contributes to throttling because now the heat from every other part of the Pi forms a hot thermal blanket over the CPU.

I’m amazed they didn’t at least add a few passive ventilation holes or slots, somewhere. It gets crazy hot inside the case.

I'm sad about the case too, it needs some extensive Dremel work to allow prolonged database queries or movie transcoding or basically anything I'd like to use it for
Yeah snagging the official case was a big let down imo. Having to modify it just so the board will operate for a generic use case was a sad moment for me. Glad to see your solution though, time to bust out my old hole saw kit!

Also excited to follow your Dramble project!

Your phone will do this too, that's why sandwiching icepacks between your iPhone makes it do an iCloud backup faster.

Lots of vendors disable thermal limiting during benchmarks to get better numbers.

Part of me is wondering if that's why they disabled the 60C soft temperature limit for the pi 4...
Yeah, I'm kinda wondering just what the Pi is supposed to be at this point. Is it still an inexpensive low power device for teaching basic computer science? Because it seems like it's more for making Kodi boxes or the worlds most powerful LED blinkers.
I'd love to see some kind of reasonable data on the breakdown of how many Pis actually end up in education environments vs RetroPie/Kodi/PiHole/whatever else "service on an SD card image" turnkey solution type thing people use them for. At this point, my guess/gut feel is that education is an increasingly small part of the market for the Pi, yet appears to still be the main focus of the Pi Foundation.

I've always felt the educational focus was much more about the rose-tinted memories of parents who grew up in the homebrew computing generation in the UK. It feels much more a product for me in that regard than it does my kids. Ebon Upton et al are of this generation as well. British computer people of a certain age love to remember their ZX Spectrums etc.

The Pi4 is overpowered for a Kodi box.

That said, there is a huge number of applications for overpowered LED blinkers, and no lack of people to use them. And there is a great deal of educational value on having those boards available and easy to get.

Besides, it's becoming hard to find people without access to a machine where they can learn basic computer science. So this one niche is closing down, while the Pi is still unbeatable on hardware hacking.

Could you elaborate on why you think Pi 4 is "overpowered" for Kodi?

From what I've seen, there are still some kinks playing 4K HEVC videos. So hopefully, when software/firmware/etc. catches up, it should be "just powerful enough" for a Kodi box, all because of hardware decode.

Barely handling 4K does not scream "overpowered for a media center" to me.

(comment deleted)
Ok, I didn't think about 4k videos, or software codecs.
> The Pi4 is overpowered for a Kodi box.

Even for software decoding of codecs or containers unsupported by the hardware acceleration? If that's true, I might have to get one.

Why does it need to be either/or? The way I look at it, every single hobbyist, retropie, dust collector, etc is just enabling the scale that allows them to deliver the hardware at the price point.

If 100% of Pi went into schools, the foundation would likely have dried up 4 years ago. We might not be the target market, but we enable it. We enable the development, the third-party/after-market, we provide the community. The foundation provide the "noble goal", and we provide the cashflow.

Because the two goals are somewhat incompatible. The more powerful they make the hardware to please one crowd the less they can meet the "low power", and to a lesser extent price (need a fan now, etc), goals.
I can't agree with that conclusion. Where has the "low power" goal come from (as if 15W wasn't low power anyway). And as has been said already, you 100% do not need to actively cool a RPi 4 to teach kids to code with it!
> I can't agree with that conclusion. Where has the "low power" goal come from

...ok, checking again I may have read that into the goals when it wasn't.

I'm not sure low power was ever a goal. They've chosen USB as the power source, for whatever reason, and fairly consistently managed to just scrape (or just miss) the power budget. I mean, this is a device with no power management (other than temperature throttling), no sleep states, etc. If low power was the goal, they've managed to put out a successful product in spite of themselves.

It feels to me like they've only ever had two solid targets. One is the price-point, and they've shed everything that stood, and the other has been a rather solid attachment to backwards compatibility. (It often feels like the model A only exists because their original claim a $25 computer, and the A means they technically stuck it - despite it being one of their less popular models).

I think another commenter hit the nail on the head though - the whole thing feels like an emotional attachment to the way computing was learnt in bedrooms in the late 80s / early 90s - especially the success of the BBC micros (which I believe the model A & B are named after). Hooking up turtle bots, sticking wires straight into parallel ports, makes mail merges feel like a hollow shell of computing. I think that's what the Pi is trying to bring back (in a manner that makes the computer itself cheap enough to be disposable, rather than some expensive relic that you're afraid to mess with).

(Side ramble: I learnt computing (at school) in the UK in the era directly after this. As strange as this will likely sound to anyone who isn't British, the era when various supermarkets kindly volunteered to replace all our beebs with nice new Windows PCs. We went from wiring weird and wonderful things into parallel ports, to seeing computers as these expensive things that "we" had worked hard for. They went from being tools to being appliances. They weren't to me messed with, modified and tortured - they were to run Claris and Publisher, and later Netscape. My work now owes more to replacing burnt out serial controllers in Amigas, than to anything I learnt at "high school" level computing. We spent a decade or two insulating students from any nuts & bolts understanding, and modern environments are getting worse, not better. So I see Arduino, Raspberry Pi, as the antidote to being taught "computing" on an iPad.)

Raspberry Pi tower cooler -- the cooler we deserve.

https://www.seeedstudio.com/ICE-Tower-CPU-Cooling-Fan-for-Ra...

well yeah, this thing pulls 3 amps.

even though it's mostly to push amperage down the USB 3.0 ports, those electrons going across the board are gonna generate some heat.

Worst case (headless, at least) with a USB SSD powered off the bus, I was pulling 1.8 amps. But I can imagine if you’re also driving some other USB devices, and one or two HDMI displays, 3A is not far off!
Has anyone made an RPi4 lava lamp yet? Extra credit if there is a camera built in that feeds into the camera connector that supplements the RNG.
lava lamp / hw crypto module :'D match made in heaven
Just a bit off-topic but what are you guys doing that you need to do CPU throttling? I just got a Pi4 and wondering if I'm missing out on some cool projects I can do with it :)
The Pi 4 is the first generation that's actually not horrible at running Kubernetes, so I'm mostly having fun with that (ongoing saga at www.pidramble.com).
Is this just about how fast the CPU is, or is there some other reason the k8s experience is better now?
I'm sure 4GB memory helps a lot.
This. Even 2GB is a world of difference.

One of the major issues I had was the master (control) node would start getting a little weird sometimes, and it was always due to memory pressure (even if not scheduling pods on the master Pi).

Kubernetes docs _say_ 1 GB is the minimum memory requirement, but 2 or 4 GB is more realistic, because at 1 GB and no swap you have precious little overhead.

Are you planning on adding a fan to the cluster?
I’m using an open case and PoE HATs which have built in fans powered by the HAT itself.
We finally have enough performance that we can waste it with more layers. Hooray!
Home media server. Encoding 1080p video brings the temp up.
> If you run computing hardware at its thermal capacity for long periods of time, this will cause more wear on the parts than if they are run well inside their limits.

So I have a moderately increased likelihood of, at some indefinite point in the future, having to spend a whole $35 to replace the Raspberry Pi. I can prevent this by spending extra money (and time) up front.

In some cases, it might be worth it. Maybe I'm using this Pi to control something, so downtime is bad. (But even then, fans have moving parts and a high failure rate, so plan to monitor the fan's condition and be prepared to replace it.)

In other cases, if the performance isn't important to you, it may make more sense to just accept a shorter lifespan. By the time it breaks, the next generation Pi may be out anyway. I don't love thinking of hardware as disposable, but at this price point, it may be smarter economically.

I would accept the shorter lifespan just because of the fan noise. Small and cheap fans are very noticeable in otherwise quiet rooms. I'd hate it.
(comment deleted)
I hate fan noise too. Sometimes it's necessary, though. But yeah, noise is another trade-off to consider.

Regarding Raspberry Pi specifically, larger, slow-moving fans tend to be quieter than smaller, fast-moving fans. The Pi is compact, so a small fan would be the natural choice.

Big passive heatsink is not an option?
I went with a small heatsink for my Pi 3’s.
> If you run computing hardware at its thermal capacity for long periods of time, this will cause more wear on the parts than if they are run well inside their limits.

How true is this actually? Personal anecdote is an I5-2500K overclocked to 4.6ghz with budget aftermarket cooler (its at 4.6ghz constantly regardless of load). It's been running 24/7 in my desktop since 2011.

When will it fail? I see these disclaimers all the time and they sound logical, but does anyone have some numbers on these fail rates under these 'prolonged extreme loads'?

I've been running a I7-6700K on a 120mm AIO radiator for three years. It gets loud, but it also smashes my OC results.

Personally, my biggest worry is the pump failing, and not the fans.

I used AIO for many computers, but every single system failed. It was the pump every time.

I switched to a Noctua passive cooler that I've used throughout motherboard generations (Noctua always releases a cheap kit to attach it to the latest socket type) and it has never been a problem. Thermals aren't a lot worse; I used it on 200W TDP 10 core SKUs, and on normal consumer chips. It has never let me down. AIO seems like a total waste to me; lots of moving parts to die on Saturday morning and make your computer unusable all weekend.

My AIO has been running fine- but I'm seriously considering a beefy heatsink next go-around.
Generally the time under load is the factor that kills parts, not just time on in general. For example: components that are constantly used for test benches to run benchmarks that pin them to 100% performance constantly will burn out much faster than your case.

Also, CPUs can get "binned" where you get the best CPUs in a batch which usually result in increased longevity and higher ceilings on overclocking. People usually call this the "Silicon Lottery". You may have just gotten lucky in your case and won the Silicon Lottery. That's honestly an impressive overclock for that CPU.

I've personally never had a component burn out before I want to upgrade it either, so I'd take that common "disclaimer" with a grain of salt. I think it really comes down to if you got a good bin or a bad bin part with any component.

Last time I checked, a consumer grade CPU has expected lifespan of 20-30 years, so your CPU is still in its young adult years
The rough date you checked will be relevant - as CPUs move to smaller process nodes, they become more susceptible to this type of failure.
And in my experience, it's usually not the CPU itself that has an issue, but some flaky solder joint, or the BGA under a chip, just like in the Xbox 360 'red ring of death' era. With mass-manufactured products, if you got a device from a batch that was a few percentage points out of tolerance, the thermal stress can cause issues more quickly than if you didn't introduce the thermal stress.
> How true is this actually?

There's a nice write-up here[1]. To summarize, using the Arrhenius equation as an approximation, one roughly gets that a 10C increase in operating temperature leads to cutting the expected lifetime in half.

However this is a rough approximation and does not consider various other failure modes, such as those associated with thermal cycling.

It also does not really matter much when significantly below the rated max temps: 2000 years cut in half is still plenty...

The take-away seems to be that operating continuously near the max rated temperature can have a significant effect on expected lifetime. Also extreme thermal cycling is probably not a good idea either.

[1]: https://www.electronics-cooling.com/2017/08/10c-increase-tem...

Probably in 20 years. The rasperry pi will probably fail in 100 years. Checkmate fan.
I think if you focus too much on "$35" you will be very disappointed with the rpi. You need a power supply, that's $10. You need an SD card, that's $10. You need a case, that's $10. You need a micro HDMI to HDMI cable or two, that's $10. You need a fan, that's $10.

If you mentally decide "the Raspberry Pi is a $100 computer all-in" you will be much happier. You won't buy one to sit in a drawer (as many on HN have complained about), and you won't try to use a USB charger from your phone from 1992 and that SD card you got with your sandwich at IKEA and be disappointed that it's not very reliable.

I have one in a drawr, like many. At my a few relatives I have PiHole running and they love it just had to walk my grandma through occasionally turning the pi off and on (pull the power, reconnect).

But I have the current Raspbian distro in a VM for messing around with as I don't need the hardware up. Once I have a use for the pi in my drawer it's going to be useful

but when the chip fails from heat stress, you only need to replace that. The case, cables, and power supply will all presumably be fine.

$100 upfront, $35 to replace

Despite the fact that doesn't add up, it depends what you are doing whether you need all that. I've had great value from my £34 Pi 4 since I don't need HDMI cables, cases or fans. I did also already have the other stuff too but I take your point there's no avoiding an SD card (yet).
If you mentally decide "the Raspberry Pi is a $100 computer all-in" you will be much happier. You won't buy one to sit in a drawer (as many on HN have complained about), and you won't try to use a USB charger from your phone from 1992 and that SD card you got with your sandwich at IKEA and be disappointed that it's not very reliable.

It's a bit of a shame that what started out as an educational computer is about $100 to get started properly. Nowadays I recommend Micro:Bit for tinkering (our daughter has one, she likes it very much). It's about 20 Euro and you have everything needed (including a battery pack if you want to use it portably).

If you want a machine that fore serious work (NAS, media center, low-end desktop), before spending $100 one should at least consider spending a bit more and get a NUC. You can buy a NUC for around 120 Euro, then you have to add memory and storage. But at least, you get a machine with fast I/O, faster CPU, and is more expandable. It also does supports 4K and h.265 decoding like the Pi 4.

(comment deleted)
most components should do fine under these temperatures no? on PC usually VRM's are heat sensitive, but the CPU can take pretty crazy temperatures itself.

is the issue similar on PI? What component is at risk at these temperatures? 60-70 seems a little hot for a small device , but is it really TOO hot?

Damaging temperatures are generally between 90-100°C. Throttling can happen at lower temps and those are usually CPU dependent (and cooler dependent). Also CPUs with less cores handle thermal throttling worse as they have less cores to fallback on if one overheats.
It's very dependent on the device. Enterprise/networking-class chips I work with are usually rated for long-term reliability up to 125C, more recently that's come down a little bit (particularly regarding memories) but still significantly over 100C.
As a sometime hardware engineer, my rule of...er...thumb was that if it was difficult to keep your finger on the device for 5 seconds then it was overheating.

Damage to devices certainly doesn't happen at 100C but you might be talking about case temp under some assumptions about how that relates to junction temp for some specific setup. In general its junction temp that decides if the device suffers permanent damage.

Well I guess not damage, but most mobos have an overheat prevention mechanism that will shut the system down at those temps. Can't really use your CPU over 100°C if your motherboard shuts the system off.

Sure you can override these in the BIOS but then you're risking damaging the VRMs which might be rated at lower temps.

What about a bigger heatsink?
That has been tried to great success with the rpi3, so I suspect it will at least make it less prone to throttling.

Edit: I just stacked a bunch of 10sec coins on the processor (without any thermal compound) brought the idle temperature down by 3 degrees C.

Edit2: Put an carbon tablet pipe on it, filled with coins. It is not up to heat yet, but right now it is idling at about 45C. It would probably be better with some fluid in it. I will probably keep this as my cooler.

Edit3: Seriously: I have wonderful thermal numbers. sysbench for 240 seconds, temperature 62C.

This is a rpi4 with a poe hat, so my passive cooling options are limited.

What is 'carbon tablet pipe'? I tried to google it, but nothing useful turns up.
Quote:

  the metal casing helps
  spread that heat around
It does, to a degree. But that is absolutely not what we are seeing in that IR picture. Shiny bare metal is a mirror at the wavelengths that thermal camera uses. So what you see on thermal image where wifi-module shield and CPU are is reflection of ambient (room) where this picture was taken. Try waving your hand around, you'll see it reflecting there too.

To measure real metal surface temperature by IR, you have to paint that metal (ideally, with black matte paint), or apply similarly-textured sticker.

How was that accomplished? Extremely nice image!
Many IR cameras have actually two cameras, a normal optical one with high resolution (~10^3 x 10^3), and the IR one (microbolometer) which is low resolution (~10^2 x 10^2). They ensemble the two pictures later and produce these results.
I have the cheaper Seek imager, which just has a low res IR camera. Can’t afford the fancy tools where not needed! I mostly use it to check the house for air leaks and thermal issues.
That image is still highly misleading, you can see from the overlayed optical image that the CPU is not coated.

Ideally you'd want to coat any metal surface with black epoxy, and set the IR camera Emissivity Coefficient to 0.9 (the coefficient for black epoxy).

Why would the overlay need to be with the board that had a coating? Buy two boards, cover the parts on one board that are silver, shoot IR. Use second board for visible wavelength photo, combine for a beautiful overlay result.
Why would you assume that they are different boards?
It's not an assumption, it's a suggestion for making an thermal/visible light overlay photo where the IR photo shows only emitted light.
Wouldn't that change the heat dissipation characteristics and skew the results? Is a metal surface coated with epoxy exactly the same (from transmission, convection and radiation point of views) as a naked metal surface?
Can't you just put it on a thumper?

I dunno Atmos or something... ... ...

Phonons and dat shiznit!

- DAMNED ENGINEERS

:)

True, I worded it a bit wrong. But the metal case is a huge improvement over the plastic die package on the Pi 3 B and earlier—a heat sink helps a little but you don't yet _need_ one if you use a fan, since the metal is much better at dispersing the heat.
Indeed - there is enough aluminum/copper in the SoC's metal lid and PCB's inner (ground/power) layers, that the whole thing is effectively a 2-3 square-inch heat-sink. Eben himself has been talking about this technique around the time RPi3+ was released.

And your point here is completely understandable too - there is a kinda-heatsink there already. Just add a fan, and we're good.

My point is - seeing a large black hole where SoC goes (color corresponding to +20-ish degrees C by the chart) is utterly confusing, since SoC is the hottest point actually.

How does it radiate enough heat with so little surface area (in comparison to an extruded aluminum heat sink with fins?)
There is more area than meets the eye - a large chunk of PCB itself is a "heatsink" thanks to sheer amount of metal in there.
(comment deleted)
To what degree does scanning it in a cold room help?
Yep, a lot of people nowadays is using IR cameras without any calibration or understanding of the physics behind the measurement. The truth is, the camera requires an emissivity coefficient to associate the received radiation with the emitter body temperature. There are tables for different types of materials, but the best is to calibrate against a surface thermocouple or similar.

A corollary of this is, when you see an IR picture of a product with very different surfaces like rubber, shiny metal, matte paint, plastic, glass, etc. you know that almost for sure the measurement is unreliable because at most they calibrated the camera for the emissivity coefficient of one of the surfaces. And they vary quite a lot...

>people nowadays is using IR cameras without any calibration or understanding of the physics behind the measurement

The four factors are Emissivity (ε), Absorptivity (α), reflectivity (ρ) and transmissivity (t). Emissivity is only one facet. Ideal is high Emissivity and low reflectivity.

https://en.wikipedia.org/wiki/Radiation_properties

Certainly, in most cameras the factors are in fact accounted for by setting the emissivity factor, a reflected apparent temperature, the distance to the object and the relative humidity. Thermography is certainly not an straightforward "point and click" process...
As a once-infrared-astronomer, I can certainly confirm your statement: Thermography is certainly not a straightforward "point-and-click" process.

Six years of grad school...

Sounds like a cool profession, what are you up to nowadays?
Damn, are you the Cliff Stoll? I must say I thoroughly enjoyed your book back in the day. Upon reflection, it may have even affected my career (for the best).

Cheers!

Emissivity is mostly what matters in this case.

Whether the material is reflective or transmissive just means you'll be "seeing" IR sources reflected off the target or from behind the target, respectively. You do have to keep that in mind if there's an object not in ambient temperature in those locations (e.g. a clear sky is of very low temperature, a person is around 35C).

Absorptivity is generally equal to emissivity (except for things that have wildly different emissivity in their blackbody IR spectrum vs the ambient IR spectrum).
Indeed, but note it usually works well enough, because most surfaces have high IR emissivity (in particular with organics like plastics, papers, woods, fabrics, etc.; soils and liquids as well).

The one exception you have to be careful is with unpainted metal (high reflectivity). If polished it can be so low that even calibration doesn't help -- because you'd need extremely precise calibration that isn't practical and because reflections and noise/other heat sources will pollute you readings. In that case a simple sticker (or thermocouple reading) would be more adequate.

(comment deleted)
He's measuring the temp of the board next to the CPU, not the shiny IHS on the CPU.
Physicist here, and I disagree. Where else could that radiation with ~10 micron wavelength at that intensity with that localized spatial profile from that particular direction be coming from?

Yes environment typically will have some residual "noise" at those wavelengths, which you can check its intensity and spatial profile by taking a "dark frame" if you're in a strange environment and are really suspicious, but it's hardly going to alter what you're seeing in any qualitative way. Assuming someone isn't sending a focused beam of exactly that size at exactly that spot at an exactly correct angle at that particular wavelength.

Physicist here too. What exactly do you disagree with? The parent comment is sound – thermal imaging cameras typically under-read on shiny metal surfaces. Their emissivity/absorptivity at relevant wavelengths is low, and reflectivity is high. Thus, their own Planck spectrum is (approximately) scaled down by their emissivity, and consequently the radiation in the measured MIR band is mostly what is reflected, which tends to come from the room-temperature environment.

A polished piece of metal makes a shitty black body. This is also why shiny metal (foil) is used to curb unwanted radiated heat transfer everywhere from thermos flasks and cryostats to space probes. (The lower emissivity further improves the efficiency of multi-layer insulation.)

Let me try again. Assistant Professor of Physics here (not a grad student).

Yes, reflectance of room temperature aluminum at those wavelengths is pretty good (not true for all metals BTW). Yes, this usually makes it hard to distinguish thermal radiation and reflected radiation with metals. What are you trying to say though? That whatever comes off from a metal must always be a reflection coming from somewhere else?

> Thus, their own Planck spectrum is (approximately) scaled down by their emissivity, and consequently the radiation in the measured MIR band is mostly what is reflected, which tends to come from the room-temperature environment.

I don't know what you mean by "Planck spectrum is (approximately) scaled down" (as "Planck spectrum" only refers to thermal radiation and is generated in a separate process from reflected photons [one is governed by the conduction band whereas the other is governed by everything up to Fermi level] and you can't hope to suppress thermal radiation by simply shining random environmental light on a metal --there is no such thing as "scaling down" of thermal radiation unless you engineer such property), but there is just no way that 10 micron photons at that intensity could be coming from a room-temperature environment.

So your blanket statements about metals aside, the hot area in that picture is due to a very specific signal which can't be due to something that's reflected from the environment. No significant fraction of those 10 micron photons coming off from that localized the area around the CPU could have originated from the environment --assuming that those pictures aren't taken in a hot oven and someone focused the thermal radiation on to the heatsink to get that amount of intensity.

And as I mentioned, that's pretty trivial to test. If those 10 micron photons were coming from the environment as you or the parent comment suggest, the thermal camera would report ~60C even when you look at Pi 4 when it is cooled (again, this is something can use as "dark frame" and subtract off from all readings if you're trying to be more accurate). This is clearly not the case, though, as you can see in the video on the blog post.

Genuinely curious person here. What is the key reason to use photon size instead of wavelenght? To stress the quantifiableness of the radiation being capured by the thermal camera? To be honest it is the first time i've seen photon size semicasually mentioned in a conversation. Then again i only had physics in high school.
> What is the key reason to use photon size

Photon size wasn't used. Micron is an unofficial name for 1e-6 m, the lengths of 0.7e-6 m to 1e-3 m correspond to the wavelength of infrared radiation:

https://en.wikipedia.org/wiki/Infrared

So "those 10 micron photons" there mean "the photons of the radiation with the wavelength of 1e-6 m."

Ah! Makes so much more sense, thanks!
They're the same thing: de Broglie wavelength would be the closest thing if you want to assign an effective "size" to a particle.
That actually makes a lot of sense after a bit of Wikipediaing. Thank you for broadening my knowledge.

  What are you trying to say though?
In the first image in the linked article, the thermal camera picture has a scale at the bottom. On the scale shown white and red are hottest (66°C) and blue and black are coldest (23°C). The CPU is black (23°C), and the PCB directly adjacent to it is white (66°C).

kees99 and klickverbot are saying it's unlikely the CPU is actually 23°C, especially given the author's statement the CPU was around 60°C, and that it's well known taking thermal camera images of things with different emissivities will produce inaccurate results.

kees99 is also saying, given that the thermal image doesn't accurately measure the temperature of the CPU, the article's statement that the metal casing helps isn't really warranted.

CPU is at 23C? Are we looking at the same image?

The CPU is the heat generator there, and is in contact with metallic regions around 60C (the red ring, if you compare to the real picture and follow the metallic bevels), where heat conductivity abruptly drops, which is what I've been talking about from the beginning. Since the heat is generated by the CPU and flows to the metal casing and to the PCB, the CPU can't be lower than 60C.

I agree that the reading for the inner region of the metal casing (which is not the CPU) must be off, and it's probably because the emission intensity there isn't strong enough and the camera software is mixing the emission and reflection when inferring the temperature (which gives physically incorrect results because the spectrum won't obey Planck's law, but the error depends on how different the temperatures are, and gets much stronger as they drift apart) rather than doing something like a "dark frame" subtraction (which is doable in principle).

Accuracy concerns aside, though, everything we see there (when you consider the physical context) supports the fact that the metal casing helps spreading the heat (which is obvious, it's a material which high heat conductivity, and there wouldn't be any need to put it there otherwise).

Even the 60C reading must be off by some for the same reason (given the regions appearing at around 70C), of course, but I assume OP doesn't care about that level of accuracy.

I agree it's true metal conducts heat better than plastic, and that a metal package is a conventional choice for that reason.

I disagree that the thermal image provides evidence of those truths

Does the image prove the CPU has a low temperature? No, the image reports the temperature inaccurately. Does the image prove the package has no hotspots? No, it wouldn't show hotspots if they were there. Does the fact the PCB gets hot tell us much? Not really, you'd expect heat to conduct from the package and balls to the PCB no matter what the package was made from.

If that cap didn't spreading the heat as well, what you'd be seeing on the thermal camera would be something that glows around 60C-70C (because clearly, the camera's software can more or less resolve the room temperature reflection from 60C thermal radiation), and the color would be more or less uniform in the region above the CPU. There wouldn't be such a strong observed temperature gradient.

Which is the evidence you're looking for.

Understanding the reflectivity and emissivity of the material you are measuring is something that almost nobody accounts for outside of critical, professional applications.

Not that it makes the measurements in question any less inaccurate, but it's very common to see these problems in temperature measurements.

how about touching a thermometer to it? (granted that is one spot, but I'd think it's most precise of all.)
I cannot find a data sheet for its CPU (BCM2711B0). Is Pi4 actually an open hardware system?
None of the Pi's have been open hardware systems.
The Pi4 does not need a fan. There is a grand total of 10 watts of heat to dissipate, 15 watts max supply but that includes power sent to peripherals and wifi. A matchbox sized heatspreader which just touches each of the warm components will work assuming a bit of ventilation is arranged to let 40-50C air float away. The heatspreader shouldn't even benefit significantly from being a top conductive alloy - a bit of tin lid will do the job.

Experiences with shifting 200 watts for cpu cooling has people thinking dedicated alloy sinks and fins and fans are appropriate, but this is just 10 watts - it's child's play ;/

I think 'need' is a pretty strong word in this context. 'Could use' would be a better term, as the Pi 4 clearly works without a fan, just not at peak performance for more than a few minutes.

For many use-cases, I would prefer some performance reduction over the noise generated by a fan.

Yes, if you're committed to a silent use-case, your options are essentially between a processor that doesn't throttle, or a processor that starts faster, and then eventually throttles to a steady-state matching the non-throttling processor in performance.
My problem is I use a shield on top of the PI for servo connections. I have found that a PI 3 can get hot without a case but with a board on top of it. I am glad they are improving the PI but I need to think about side mounted cooling.
I'm happy I have not bought the Pi4 yet, fanless for this kind of boards is the bottom line for me, if Pi4 needs a fan, then it crosses the line at least for me.
Reading all those IR images popping about the raspberry made me wonder why we're still using a "rainbow" (HSV) colormap.

I think there are better alternatives, that have both a linear luminance variation, and are colorblind-friendly:

https://news.ycombinator.com/item?id=17723041

Raspberry Pi is becoming like a NUC. Meanwhile, there are people using Intel NUCs as fanless Hackintoshes:

https://www.youtube.com/watch?v=tUUP8K3RqAo

Except the Raspberry Pi is an ARM SOC (ARM64 in this case), while Intel NUCs are x86-64. macOS requires x86-64 for now. If Apple ports it to ARM, I really doubt they'll make it easy to run on other ARM hardware.
They sell cases with heatsinks and fans for $10-$20. I would just recommend buying one of those versus modding one like the guy in the article did.
I feel like if I buy official parts, they shouldn't require modding to get the max performance out of the product. They should work without issue together. Currently you can barely get moderate performance with the official case.

This article and mod are helpful, even if obvious, for people like me who bought an official case with my RPi4 expecting them to work well together.

But they do work without issue, until you start stressing the CPU. In the majority of normal workloads, you don't run into thermal throttling, especially if not using the official case.

Also, if you have the official case and want to do this mod - for a few bucks more than the cost of two "pi fans" (which run $8 on Amazon), you can get a well designed case with heat sinks and a fan.

I see very little reason to actually do this mod versus buying a properly designed case besides the urge to tinker.