Probing Pi 5 silicon with an Electron Microscope!
Vložit
- čas přidán 4. 06. 2024
- Visit brilliant.org/JeffGeerling to start your 30-day free trial, and get 20% off your annual subscription!
The Raspberry Pi 5 can't be overclocked past 3 GHz. Why? The secret is hidden in the silicon deep inside.
X-rays don't go deep enough. This video explores modern CPU architecture, clocks and PLLs (Phase-Locked Loops), silicon die shots, and even transistor-level debugging with an electron microscope!
Mentioned in this video:
- Blog post with more images: www.jeffgeerling.com/blog/202...
- Open Circuits book: amzn.to/3uZRJow
- Kleindiek Nanotechnik: www.nanotechnik.com/copyright...
- John McMaster (Twitter): / johndmcmaster
- Siliconpr0n die shot - RP1: siliconpr0n.org/map/raspberry...
- Siliconpr0n die shot - BCM2712: siliconpr0n.org/map/broadcom/...
- der8auer's video on 7nm analysis: • Unreal Precision - Ana...
- Water cooling block for Pi 5: www.jeffgeerling.com/blog/202...
- EDATEC's passive heatsink for Pi 5: www.edatec.cn/en/ac/pi5-openc...
- Apple II Pendulum Clock: github.com/wkjagt/apple2_pend...
- 32nd note drummer: • 110 120 Bpm 16th Note ...
- Monster vs Dave Grohl drum battle: • Dave Grohl and Animal ...
Support me on Patreon: / geerlingguy
Sponsor me on GitHub: github.com/sponsors/geerlingguy
Merch: redshirtjeff.com
2nd Channel: / geerlingengineering
Contents:
00:00 - Human DNA on my chip?
00:42 - Overclocking, featuring actual clocks
05:59 - Heat is not an issue (unless you're a candle)
07:14 - Die shots: OCP on the Pi 5's BCM2712
11:51 - Die shots: RP1
12:37 - Probing individual transistors at 16nm
18:35 - Peering deeper inside silicon
20:06 - Learning more... - Věda a technologie
Magic smoke is what I keep in my computers
I did not want to go *that* deep. That's hiding just under the silicon layer, and if we probed too deep, it would escape, and I would not be able to show you this video :(
@@JeffGeerling good idea, would be dangerous to reveal the secrets of the magic smoke
As long as the smoke stays inside you should be fine. 👍
ahh i have a magic keyboard and mouse... i can accidentally (seriously!) kill a server at the opposite side of the city im in! im dangerous but very curious... the bits that live with the magic smoke are ultra secret!
im fascinated by this! its fascinating!!
I thought the smoke was just a byproduct of the Imp getting annoyed and portaling out of there 😊
Inside your computer are millions of tiny bees doing math, so you don't have to!
That's why I regularly run a program that does nothing more than just count up. The philosophical humor of "seeing" electrons do literally nothing of importance at our whim.
remember at the end of the wicker man when they stuck nick cage's head into a computer?
Millions of nerds you say?
As a U.S Navy trained electronics technician, I've been around long enough to see magnetic amplifiers in action, but it blew me away to see those inductors on the metal layer. We've come a long way in 40 years. Nice job Jeff.
The crazy things they can build up in the space of a few mm^2 is mind-blowing.
And yet, these inductors are much bigger than the transistors
A friend and I swapped a clock crystal with one from a TV in a Commodore 64 in the 1980’s. Lode Runner never ran so fast. Then the computer made magic smoke. Cooling wasn’t something we considered until it was too late
Hehe, sounds like the first time I tried running an older AMD cpu without the little heatsink it came with. It was fine until it wasn't, and a pop and a smell was heard! They didn't have the same kind of thermal protections in the chips back then!
@@JeffGeerlinggood old skt 462 Athlon’s. Zero thermal sensors and absolutely zero thermal protection!
Magic Smoke is an electronic engineer's nightmare.
This is amazing. From 1996-1999 I worked for a company that reverse-engineered ICs. I wrote software to help image the chips and assist with their reverse-engineering. But the state of the art in 1999 was 140nm. So smaller than visible light, but not all that much smaller and it was still practical to use optical microscopes even to see individual transistors.
Of course, our reverse-engineering was destructive because we took photos of each chip layer (top metal, next metal, etc) and then etched it off to clearly see the next layer.
I recall one project, however, where they contracted someone with a focused ion beam machine to deposit new metal wires on a broken chip to repair access to onboard flash memory that they could then read out. Even at the time, it seemed science-fictiony.
Those probes are amazing. They go from a very much human scale too all the way down to nanometer sized tip
Yeah; I actually asked how they make them-it's a third party that makes these probes, but the way they're made is pretty resilient. I would've thought they could break just by being handled roughly, but apparently they are pretty resilient (all things considered), unless you intentionally bend one!
@@JeffGeerling They're etched. You can actually make those probes at home. Yes, really! They are, ironically, the easiest part of this whole setup. ~100nm tips are possible using a bench power supply and some kitchenware, going below requires some prep, but it's all just chemical/electrolytic etching.
The multi-axis nanopositioners, on the other hand, are quite complicated.
@Spirit532 are there any videos that you can point us to that show the way these probes are made?
@@mr.davemaeen8136 Look up "tungsten probe etching", lots of papers and articles on that.
These deep dives into chips are truly amazing and the people who design them are even more amazing. I remember in the 90's having a pleasant chat with a chip designer and was open-mouthed. What was done back then was nothing next to today's capabilities. Back then I was really impressed....today I don't even know what term to use.
It was in some ways harder back then, without some of the modern tools and much slower ones that did exist. But complexity grows so as tools get better designs also scale up!
Watching that electron microscope come into focus was magical. Insane what tools we have.
Close to magic, but I'm also amazed these bits of machinery were all designed like 50-80 years ago, they're just better today but not fundamentally different.
@@JeffGeerling Wow. I would've thought that it originated in the 2000's.
@@JeffGeerlingnope, not fundamentally different, but phenomenally better.
Back when I was in school, way back during the ice age, my junior high and high schools had donated transmission electron microscopes. The most common one we had was mid-60's vintage, with a early '70's vintage as our flagship TEM.
While they might, just might've had the resolution to see most of the features on these chips, they'd also have destroyed the chip, due to their higher beam current.
Alas, by the time our kids went to the same schools, those electron and even the optical microscopes were gone. High school chemistry class no longer used reagents, they literally used jars of M&M's.
No wonder kids get out of school and need a minimum of a year of remedial teaching in college today!
@@spvillano It's amazing how incremental improvements stack up over decades. Just like one generation to the next of something like the iPhone wasn't always that noticeable-try the original vs a new iPhone today and there's not even the same league, in terms of speed, sound, screen, battery life... everything! (Despite the iPhone being a slightly worse example than microscopy).
@@JeffGeerlingevery new generation of PC, the first thing I do is slap a VM on it and load DOS and Win 3.11. It's been years now that the full GUI is up in the amount of time a POST test used to take.
As for electron microscopes, the improvements are vast. Our old 1960's models had x-ray hazard warnings if the beam current was set to the higher settings (which would burn through the specimen grid anyway), now, they use a fraction of the beam current for higher quality magnification and imaging.
I seriously believe that my grandkids will likely be using stuff straight out of Star Trek by the time they're in their 30's.
As a factorio addict this video was a trip!
What's inside my computer are electrons doing their thing needed to play another amazing Jeff Geerling video.
Just to be clear: the "15nm" doesn't actually refer to the size of anything. Manufacturers have long given up the original meaning of "feature size" and now it's just a marketing label.
Since you are actually visualizing individual transistors, it would be interesting to make a measurement in the micrograph to see what are the actual dimensions of normal transistors.
They do vary in size, though, with some being much longer in one direction.
If you want to go down the weird PLL stuff rabbit hole: look into
Clock Domain Crossing (CDC) theory, metastability and the electronic version of Buridan's principle. Happy Pi Day, everybody!
I still remember when PLL's used to drive many consumer electronics tech crazy trying to figure them out. No mystery, one just needed to actually understand the theory of operation.
Later, saw much the same concept barrier lay up many a tech over SMPS units, similar failure to comprehend the necessity of precise feedback.
Your explanation of clocks reinforced the (joke?) that we tricked a rock into thinking with computers. :) Happy Pi Day too!
Replacing the oscillator was the first thing that came to my mind, I'm glad you covered why this probably wouldn't work. I've also been quite impressed by the performance of the standard Pi 5 active cooler, it's even kept my overclocked Pi 5 well under thermal throttling. Fascinating video!
EM is pretty freaking amazing. I’m a biochemist and we can study proteins down to about 6A resolution in their native folded state using cryo-EM. 6A is 0.6nm and very close to the size of a single carbon-carbon bond. The cryo part is used to freeze a solution of pure proteins down to near zero kelvin to prevent any vibrations in the sample. So yeah these things are built away from trains and highways with buildings built up specifically to minimize ground vibrations.
I wonder what we’re going to do as transistors continue to shrink. I assume we will be hard limited to the atomic radius of silicon itself.
Yeah... creative packaging can still get a little more 3D, but at some point, I imagine we'll have to find a way to 'grow' circuits instead of 'print' them, to get any reliability when we get closer to atoms.
Well, Jeff, I'll tell you what… the next time you are over here, in the UK, and visiting the World's only Raspberry Pi Store once again, you MUST take a little detour to the edge of town and visit the Cambridge Museum of Technology where, in the Pye Building - a small hut out the back of the Victorian Pumping Station (which houses the museum) run and managed by the Cambridge Industrial Archaeology Group, you'll find a display explaining just how the Cambridge Scientific Instruments Company played their part in the development of those Scanning Electron Microscopes. They even have two original machines on display, although I do wonder the last time was that anyone tried to switch one on! These machines look like Mini Mainframes from back in the day. Quite amazing!
Heh, those ancient SEM's likely would vaporize those chips, if they managed to get the things to emit and begin scanning controllably.
If they ever want to light one of them up and actually have it function, they're free to call and I'll help out. I still remember how most of those circuits work, so restoration shouldn't be too hard and it'd be worth renewing my passport for. I'm old enough to have worked with both electrostatic beam handling and electromagnetic beam handling.
I'd also love to see that SEM in the modern lab's vacuum pump. I remember when plasma volatilizing organics would create a new task of changing out the oil from the pump.
That was a lot of good background information. Thank you, and those who helped you, to bring us a deeper dive into silicon. I had a good chuckle at Red Shirt Jeff getting his fire on!!
+1 for the OCP reference
Really like the die shots. The inductors at 9:44 are much more likely to be part of an oscillator in a PLL (typically in a transceiver, which requires a very stable PLL that often uses inductors) or an RF circuit, rather than filtering power. That's also why you see them round them corners since they are high speed I/O.
And metal fill is to keep the layers flat during polishing, more so than timing.
Excellent content, Jeff. Thanks for the trip down the rabbit hole.
Bro. After watching this video and then like 10 others that help explain more. It's still just magic.
Man I remember in university, playing with a Kleindiek nano probe system to measure electrical characteristics of a single 100 nm crystal. These things are so friggin cool.
At 4:19 there is a very high frequency noise, kind of hurt with headphones on
Strange, I couldn't hear it with my volume turned all the way up and my desk studio monitors-but maybe it's something outside of my hearing range? I included the audio from that pendulum clock for the Apple II, and it might just be the background noise from that clip :(
I just went back and heard it too. It wasn't painful though but I am on speakers not headphones.
@@ssgLunchbox Ah, if you have it turned up loud enough (or have *really* good tweeters and it's turned up enough), there definitely is a pretty high frequency sound coming through from that pendulum clip.
I'm not sure if I can tweak it using the limited editing tools CZcams permits.
I will have to add in an edit step for rolling off the EQ for any clips I pull in from outside sources, and sorry about your ears, yikes!
@@JeffGeerling Don't worry my ears are just sensitive to high frequencies, most people probably won't hear it, luckily they aren't bleeding :P
Heard it almost immediately myself as well. And im on 50% volume both on the video and the PC, and wearing headphones.
Not the end of the world and in fact a lot of youtube videos have this problem because the editors cant hear it themselves, whether due to age or damaged hearing.
I got human DNA on my chip one time. The "Geniuses" at the Apple Store weren't too happy about it, and couldn't work out whether it was velocity or shear volume behind exactly why it was found that deep inside the laptop.
cap
😳
🤮
What. They showed me a stack of 30 other laptops with the exact same problem, from just that week alone. I'm just a statistic here.
I love that the company is called Kliendiek (Klein = small in german)
Insightful and concise as always, Jeff. Thanks for always giving us your best and never half-assing your content. It's greatly appreciated.
@6:33 enter redshirt Jeff, flame on!
This is fascinating stuff, keep it up Jeff!
I know it's a nice analogy to say that the nm level chip structures bend light like a prism but the processes are fundamentally different. The effect in a prism (or raindrops forming a rainbow) is dispersion, where the refractive index changes with wavelength. Whereas the effect on the chip is diffraction where the physical distance between objects changes the phase of reflections, which then add constructively, or destructively. Yes they both change with wavelength but the processes are completely different, and equating the two glosses over almost the whole of 20th century physics with quantum mechanics.
It is important to try and make complex things understandable, but I think it is also very important not to drive your analogues too far. Yes a bike and and airplane both get you from a to b, but in a very different manner, the differences are probably more important than their similarities.
i bought the book awhile ago and love it jeff, thanks for the vid.
Thank you my brain has exploded.It's just amazing we gone from valve to this.
Misinformation. What's actually going on is that there's a bunch of little goblins doing math
True
Yeah, but AI's biggest advance was replacing the goblins.
Unfortunately, they replaced them with gremlins.
Then, fed them after midnight...
So cool to see behind the various curtains and get more of an understanding. Cheers to all involved in this one!
Fine job, Jeff, great explanations! And hats off to Kleindiek Nanotechnik for their contributions, as well.
Fantastic video Jeff, very informative and thoroughly fascinating, great stuff!!
Loved this video and the production quality Jeff! Great work!!
Keep up the amazing work you're doing
This is a really excellent and educational video. Thanks!
Your best video so far. Period.
Heh, I wish *CZcams* thought so.
beautiful deep dive !
17:34 voltage is measured in volts, not amps
vampires
What he said was "amperage, at a particular gate voltage". There's a relation between these two values for a transistor.
He's definitely NOT a tech and lacks basic understanding of basic electronics. He is annoying.
OLD tech Precision eq repaint, PBX and Sonnet have general class FCC license...
yes a relat5ion and how he talks is not correct. stop making excuses.@@boneappletee6416
Great job on this video Jeff. I really enjoyed this.
really cool seeing the various values and instruments used in real time on a sample. never seen that before.
I was breathless for most of the video, this is incredible. I was not even aware someone is able to do this kind of measurements ...to be able to test a single 16nm transistor is just mind blowing.
Incredible indeed. Thanks Jeff!
Jeff, this video is excellent. You clearly put a lot of effort into making this one, and it stands out in my mind as one of your best videos. Keep doing what you do!
Just a heads up there's a super grating high pitched tone at 4:19 something like crt whine with but with the gain turned up ALOT
;-;
Sorry about that-I didn't catch it in the edit, it came through from that Apple II pendulum clock clip :(
You misspelled „NDA” 😂
Heh, true!
Wish there were more open designs, even RISC-V chip manufacturers seem to not be as open with their IP as I think most of us would like.
@@JeffGeerling Its sad there's not a lot of good open source blocks.
That brilliant shot of the circuits would make a great tshirt or mug design
Another great video on Raspberry Pi 5! Thanks @jeff.
Ah, C0 may be the third silicon revision, but who knows how many metal layer revisions (the number) they went through before
Джеф, очень люблю твои видосы по raspberry pi, каждый раз рад как первый, когда выходят твои видео. Больше всего люблю моменты когда автопереводчик в моем бразуере переводит "raspberry pci" (подразумевается PCIe) как "малиновая пися"
Fantastic! Thank you.
I like your approach to talking about this subject. Some of your more obscure topics that are on this channel just take me out of what I would think is interesting about it. (not to say your point isn't getting across just that I may not enjoy the topic as much as someone else and yet some still are either too condensed or not well enough broken down for the layman) I think your vids are pretty interesting and I have watched a few. Its just that for some of the info I guess isn't for everyone (not that that is a bad thing) It may just be that I understand the context of this video better then the others. Either way I think you are doing a pretty good job. Thanks for the vid.
My favorite part of your videos was the bloopers at the end. I miss those.
It's so amazing that humans were able to do all of this.
regarding those dots.. we call it cheese and fill. Depending on the foundry and technology, it is either added by the designer or tapeout engineer (fab). It is true the help with interference, but from the manufacturing side, provides density or structure during CMP. Planarization is unevent on areas without metal.
The quartz crystal isn't the oscillator unless it is crystal oscillator module containing an amplifier and oscillator circuit along with the quartz crystal.
The two common methods of producing a clock signal for a digital circuit, not withstanding the use of PLL's are: standalone quartz crystal and separate oscillator circuit, and a crystal oscillator module which outputs a square wave at the required frequency and contains the crystal and oscillator circuit. The latter are easier to use.
The quartz crystal acts as a filter of very high Q factor which is used in the feedback path with an amplifier to create an oscillator circuit.
That very high Q factor of the crystal creates a very narrow bandwidth filter which restricts the oscillator from oscillating at any other frequency other than the frequency indicated for the crystal.
Wow! Wow! This is one of those videos that changes my perspective on the world! This is equipment is as amazing (or more) than the chips themselves!
Often the test equipment is even more fascinating, because it has to get down to accuracy that's even better than the devices they test!
It's nice to see Red Shirt Jeff escape from where ever you keep him!
So cool! Great work as always.
Fanboy here: A collaboration between you and @AppliedScience would blow my mind.
I can also recommend Branch Education's video "How do Digital and Analog Clocks Work?". This explains how the clocks in computers work in detail.
Amazing views in this video wow. Good job.
learnt a lot from this video, great content :-)
I was blown away when you said that there were 11 layers to the chip before you even get to the silicon! It's all a bit TOO amazing for me to comprehend
I-V curve of a single transistor on a Broadcom BCM2712 16nm processor; too cool. Thanks.
Now I will definitely use a blowtorch to light a candle.
you down with operating performance points? ya you know me!
As someone who assembles electron microscopes, it's nice to see the result.
And yes, electron microscopes are all about purity of everything. A few days ago, I was cleaning the chamber of an electron microscope for 12 hours straight...
Wonderful! Ok now in english!
Aww, man. I was actually much more excited to see what was hiding deep in the RP1 than in the BCM2712 hahaha.
Heh, we'll get there!
I have been building electron microscopes but never so big that buildings were built around. I used a technique to bend the beam to get its required length, sort of electrical prisms.
I love my RPi5. Even reliably stock, it's more than fast enough for Ham radio modes. OBS only feels 50% faster, and SDRs still struggle. But it runs everything else I need to without snags. I just viscerally hate the non-standard USB-C supply. 5V5A is not a thing! They should have gone with 12V3A and convert it back to 5V. I made my own 13.8V to 5V5A USB-C supply, and it still gives me the warning at startup. I added the script to bypass the warning, and it ignores it.
Quite interesting.
There is a difference in how Prisms break up light into a spectrum compared to the effect you see with silicon dies.
The effect the die produces is diffraction, which is passing light through tiny slits rather than refraction where light is bent due to the medium density the light is traveling through.
A good comparison analogy is a pin camera vs a lens camera.
Refraction is also why you get the rainbow effect with CDs and DVDs: the spacing between the concentric circles (yes, I know it is one continuous groove) are quite tiny.
Ah true. I was mostly trying to get across the concept that it's light being all crazy either way, but it's better to be precise!
No wonder it takes time to bring the chip into the market and is equally time-consuming to build software around it.
Even a "simple" chip like one of these SoCs is a 3-5 year effort, and there are so many little IP blocks inside that I'm sure some of it never even gets exposed to us end users. Like I know there are at least two or three features on the chip that aren't exposed to Linux at all!
it big task is the write the TRM for these SoC and peripherals chips.
Love the visual at 6:27
you just sold a copy of "Open Circuits"
Things I didn't know i *NEED* :P
i wonder how they keep the equipment still enough so that the microscopic vibrations through the ground of people walking and vacuum pumps running etc don't cause those ultra-thin probes to wobble. The science in making the probes themselves would be interesting to see.
I am pretty sure you know this and decided to not include it in the video, but I just did have to explain for some time to some else than the "5nm" "2nm" or whatever "18A" you want to call you lithography generation nowadays does not mean the feature are this small, since quite a while ago the naming indicate a "performance" class and not the physical size of the feature in the silicon
For the curious the transistor gate pitch of TMSC 7nm class (inclued N7 N7pN7+ and N6) is 57 nm, some feature are smaller and if you curious the Wikipedia article is very neat to nerd on en.wikipedia.org/wiki/7_nm_process
This is a good point, and you're right, I tried to keep things down to a lower level and not get into the muddy mess of what a process node is and feature size vs minimum resolution, all that fun stuff. I more wanted to explore the tiny nature of modern silicon in general, hopefully helping some people get into this miniature world and learn more themselves. I find it fascinating!
Love the 'Friggin Lasers' reference.
Best pi day present!
Happy pi day- this video is amazing
How are human’s this incredibly smart to design and create this amazing and microscopic products.
fyi, the apple ii pendulum clock clip has a high-frequency tone which you might want to keep an eye out for the production of future videos; still, incredible content!
Awesomely educational and entertaining content. You can truly claim the title king of Pi 5 overclocking. Long live the king!
Open circuit is an amazing book
Jeff, I was doing some shopping this morning (as I have been for a few months now) to eventually replace my M1 MBP 14". The battery life is stellar, I've witnessed the fans turning on maybe twice, it runs cool, and is very performant!
These are all things I would now want in a replacement. I figured I would ask the "Ambassador of ARM" himself. Do you have any experience running a Linux desktop based distro on any ARM laptops? Would love a video on your thoughts or suggestions if you have any!
Some people are running Asahi on their Macs, and the experience is good as long as you can live without some of the features :(
Otherwise, I know a few people running Ubuntu on Lenovo and have a decent experience. Otherwise, Fedora, but it seems like Ubuntu has a little more momentum still for the 'everyday' user, especially if coming from macOS.
macOS is still great though-I would consider running a Linux OS inside of UTM or something like that, so you can have the best hardware and still try out Linux on it.
@@JeffGeerling Right now, I'm actually running OrbStack. I found it while looking for a MacOS equivalent of WSL, as I like having the ability to run a couple quick commands as a sanity test. OrbStack fits the bill quite nicely!
Don't get me wrong, I like the UI on my mac well enough. I think I just want to daily a linux tiling manager for a try.
Videos such as this one make me doubt my abilities to create a contemporary CPU from available ingredients if I were to be dropped onto a desert island with no supplies or reference material
Fascinating, the stuff we take for granted these days.
EBIC and EBAC... I really thought the next thing you were going to do was break out in a song and dance about EBCDIC and ASCII too.
Quartz oscillators can't really get much much smaller, but MEMS oscillators are the next smallest.
Measuring voltage in amps? Now that's impressive!
훌륭한 컨텐츠 감사합니다!
"LTT and GN" never thought I would hear those names together given their beef.
I still have this optimistic idea that those two will make up and bury the hatchet. Last year was a weird year there, I think it could've gone a lot better and still resulted in a lot of positive change.
Fantastic!
IOW, Jeff Geerling has achieved the impossible by overclocking a Pi 5 to 3.1 GHz.
Hehe I wish solving every impossible problem were as easy as that!
If you use a Scanning electron microscope, can you see which circuits are on or off?
Jeff is now adventuring in the Microverse with this vid. 😂
Nanoverse!
Maybe I need to play a little Nanosaur, too.
@@JeffGeerlingif you meet the Micronauts remind them they still owe me 20 bucks. 😂
This gives me so much PTSD... silicon level debug.
Not sure what tools they are using for polishing, Allied polishing wheel remove material parallel to metal stacking (imaging purpose) while plasma etching with chemical batch is good from anisotropic removal. Plasma etching was over charging the IC, so there is alway a trade off
Maybe the next version will have hardware video encoding support.