Andor Gafotas - 2012-08-02
Talk in HOPE09, NYC Modern computer chips are using transistors with features as small as 22nm. They are produced in factories that are 10,000 times cleaner than an operating room that can think like Skynet. Combined, the chips they produce run everything from your cell phone to the Internet itself. While outsiders might see it as the realm of multi-billion dollar corporations, in reality, it has been achieved through a hardcore application of the hacker mindset. Each new advancement involves hacking the theories of electrical engineering, hacking waves of light, and sometimes hacking physics. In this talk, we will go over how and why the design of a modern nanoscale transistor was developed. We will also talk about the processes used to build them, and the incredible equipment that makes it all possible. Plus some fun stories about what goes wrong.
This is a cool talk. In 2019, seven years later, EUV is now actually being used in some fabs.
...and I'm watching this on a computer where CPU is on a 7nm architecture.
@TheBatracho and I'm watching this after IBM announced they have invented 2nm chips...
@Lee To be fair, neither 7nm nor 2nm actually have 7nm or 2nm transistors, respectively. It became an "as-if" measure as the actual 3D structure of the transistors and the gates they build got a bit too complicated for marketing :) Nowadays, x nm refers to a peak transistor density (yay marketing), rather than actual feature size - and even that's a marketing lie. The next time you think about why Intel's fabs are "so behind" the cutting edge in nanometers, that's part of the reason. For example, Intel's 10nm chips actually have a higher peak density than TSMC's 7nm chips, and twice as much as Samsung's 10nm chips. The IBM 2nm chip? You might expect three to nine times the density if you assumed it refers to the size of the transistors or minimum feature size or something. In reality, the peak transistor density is only 40% higher than Intel's 7nm. Just like back in the day, comparing chips from different manufacturers based on their listed characteristics (like TDP, clock rate, "nm"...) is essentially pointless. And then you learn that Intel is backporting its 10nm architectures to 14nm for... reasons.
Chip making is complicated. Even without MBAs in engineering roles and retarded marketing.
Daaang you are proper smart AF.. impressive. I should have stayed in school.. well I guess I missed the smart train.
I will never complain about CPU prices again.
witeken i wouldn't call DRAM easy. It's a different kind of challenge altogether. Yes, dram reuses cells many many times, but arguably processors do this as well.
Sure, designing a memory array and decoder from a floorplan perspective isn't as complex, and the tolerances on logic chips are somewhat stricter, the difficulty with DRAM is a manufacturing and financial challenge.
The dram industry is incredibly cut-throat and competitive, with low profit margins : as a manufacturer selling an 8GB ddr4 module, you have to fit about south of 100 BILLION transistors onto a module and sell it for less than 50$ to an end consumer, AND you have a couple watts to do it.
The memory guys are greatly under-appreciated
OK, now we have chips.
But where is the fish?
@witeken i dont follow u, where is logic in the chip or is it in the software.. not a computer guy, sorry
@Dani The heart of a computer is its so called instruction set: a set of instructions that a computer can execute, such a addition, multiplication and many others operations, basically mathematics of boolean (0 and 1). The instruction set is implemented ("hardware") on a silicon chip with transistors (which conveniently can switch between 0 and 1). So, when software (code) is written ("programming"), it is then translated to those instructions from the instruction set, which is what a chip is designed to be able to execute.
@Dani The logic is in the combination of transistors. software is also logic because it tells the hardware to form new circuits by turning on and off certain combinations of transistors.
Software is stored on a special arrangement of transistors usually called memory. Software is formatted with an instruction set and a data set.
The hardware is arranged in such a way and expects the software to be formatted accordingly. The hardware is also set up in such a way to allow the software to turn on and off certain combinations of transistor (essentially creating new circuits) in order to perform logic functions or arithmetic calculations.
The Hardware is the substrate that allows electrons to flow.
The Software is the electrons stored on and processed by that substrate.
The Logic is contained within both.
This is the practical application of alchemy.
Basically sand into gold
It was great how you said you wish you could learn how Intel makes it's newer gates and share it with everyone. You had a great presentation. Thanks for sharing/spreading your interest with us all.
This lets me see my Digital Logic textbook in a whole new light. Every individual logic gate is an impressive feat of (close enough to) perfectly replicated engineering.
It was a bit of a surprise to me just how much transistors aren't logic gates when I built my first homemade CPU from discrete transistors (ages ago). Suddenly, all those "spare" transistors in logic gate and circuit designs made sense - a "gate" that works fine on its own no longer worked in combination with another "gate". Funnily enough, all it took to start thinking about "electricity flows" rather than digital logic to make it really click - the lower level works really different from the higher level, even though they look deceptively similar (and even seem to be working, to a point!). It makes perfect sense thinking about transistors like switches, and then, when you start thinking about the flows, you realize... hey, this is flowing backwards! And these are shorted out!
the "FOUP" video was absolutely beautiful xD especially the commentary that went with it
Amazing talk! Thanks for sharing Andor! And thank you Jebnor from the EEVblog forum for bringing this to my attention. An hour well spent.
I love the hacker community. They somehow find a way to explain shit. I don't know how they do it , but we're like on the same wavelength or something.
"May 2012 Intel is doing 7nm R&D" And here in 2021 they are still chasing that 7mn dream
Will be interesting to see what comes after the transistors, just like the transistor made the vacuum tube obsolete, some new tech will soon make transistor obselete, will be insane!
What a gem. This video deserves way more upvotes. Its SAD how a drunk cat video shown stumbling on a slippery polished wood floor gets millions of likes meanwhile, a highly educated guy's hour+ lecture barely has a few thousand upvotes 🤓😲
That's what happens when common folks are 'entitled/take for granted' sophisticated technology.
Technology should be privileged for scientific minded people who are watching this video.
I now understand why Intel has struggled and is currently struggling getting their lithography down below 14nm.
Holy shit. Mad props to AMD for figuring it out! Or atleast cheating the system a bit to put only the transistors at 7nm!
It was a good talk. Only nerds like us would quibble about which particular particle is actually flipping bits. It didn't affect the talk quality in any way.
Totally superb video, excellently explained. 100%, top marks mate.
has there ever been an updated version of this presentation?
EUV is now used, so well done engineers and scientists!
It's interesting to hear about EUV from the 2012 perspective
The whole transistors not being digital thing really messes with non-electronics people. I still encounter folks that think they've been ripped off because the amplifier they bought as "100% analog" has some IC chips in it.
"All chips are analog" can really confuse people.
You're right, i noticed a few small errors in his talk. Human hair width is in the double-digit micrometer range, not nanometer. Also at 49:35 the slide reads 450nm, where he meant 450mm. Cool talk!
He says a human hair is 40 microns, what's the error?
@radeklew None, just another "know-it-all" so common in the youtube comment section. Dumb useless fuckers.
Oh now that I watch that part again, he does say 40nm just before that timestamp, my bad. I think in another part he said the right number
It's nice when you stumble across ancient history.
I just wanna casually walk into a FAB with a Rigid shop vac I just used to clean a house where they finished sanding the drywall, and let that thing rip, and then be like "aww shit, I left the filter out."
thank you for sharing ... this was helpful
That 18 month "cheaper to buy new ones" observation is probably why office grade budgets are overblown and ran by people still with AOL adresses, and constantly closing and opening under new shell titles.
I'm still waiting to hear that some big tech giant is developing product in micro-gravity.
I'm a person who likes to understand how things really work. CPU really bothers me. A modern transistor really bothers me. The things we can do with a computer bothers me - because I really cannot fully understand the 'magic' behind it eventho I've been working with computers since I was 9.
i still dont get how a transistor works, what sends the message for it to go on or off?
I know I'm replying to a year old comment, but if you're having this problem you're starting at the wrong place
Transistors are 99% of the time simplified to electrical switches because it makes understanding them much easier than starting out with "This enhancement mode MOSFET operates in saturation mode and can therefore only conduct 10A if there's a voltage difference between the gate and the source of 9 and a voltage difference between the drain and the source of 12 which it can reach within 30ns if there's only 0,1 ohms of electrical resistance, 1 pico henry in the lead and 10 femto farad of capatiance" or what have you
Point is that you should take it slow and start in an easy place like logic gates, karnaugh maps, binary etc. and then probably move onto transistor level gates and how they're actually made
Video card I'm using to watch this video uses 7 nm transistors. And they're commercially reaching 5nm transistors. Back then 12, 13 nm was still the future to come. Lord did technology advance or what.
I'd kill for a 2022 version of this.
I love how my transistor is 15 nm (or 25, i am not sure), and i think that 30 nm is too large XD.
I thought the problem with solar radiation errors was not neutrinos but muons?
Anywhere I can find the PDF of this presentation? it's kinda hard to see what's going on the screen
Silane and sarin are two completely different molecules. (19:50)
Witches used hex bags, Programmers use hex files !
Very cool talk. Thanks. At the very end: "I don't know. It scares me." What does?
(APPLAUSE)
Awesome, man.
16:05 "...Immersion lithography ... Intel's doing this now, WOW! Basically we're taking advantage of the effect you get when you try 'n throw a rock at something in a pond - it misses - [because] water has a very different index of refraction then air."
Uhhhwwhaaathefff?
17:43 "Intel just invested $1 billion getting this to work"
ASML was surely only too happy to sell them this "one off" [18.27] 'Extreme UV' Computational Litho. Dev. Rig/tool.. 18:40 "in 10, 15 hopefully 5 years or we're in trouble." I'll take Forbearing Statements for 5 or trouble, Alex. ;]
Thankyou for blowing me mind. :D
watching this on my 7nm AMD CPU, FTW!
Wow I'm watching this in 2020, and EUV it's an actual thing now! Wow...
People in the semiconductor industry always draw the OR comparison. But it’s totally misleading. Only the “sterile field” is clean (anything green or blue). Every other surface in the room is by definition contaminated. All the extra people in the room other than the surgeons (up to 3) anesthesia team (up to two) and scrub nurse, are there to eliminate contact between the two. The reason there is no reason to be that clean is sub micron particles are not infectious. Organic molecules are huge. Minimum meaningful size is Hundreds of thousands of atoms. So it’s already many times larger than a silicon channel. The vast majority of things organic do not survive in the air. To survive in the “air” said bacterium or virus needs the equivalent of steel girders inside to hold it together- much like a space capsule remaining pressurized in space. Imagine riding a random vehicle from earth into space- you would die riding 99.999% of them. Bacteria are the same. COVID probably is airborne, though the ones that survive this trip are much weaker, so airborne spread is much less common despite the greater spread area. So for most things you only have to worry about spray. Bodily fluids. If someone had something airborne then you had to take airborne precautions and then it was much more like an early fab. Bunny suits, positive pressure ORs (or negative pressure if it’s nasty). ORs are like anything else. You create them depending on your requirements but not really much beyond because the economics aren’t there. Airborne precautions were fairly rare when I was in the OR daily (pre covid I don’t know what it’s like now. )
Oh Operating Room. I was reading OR as in logical ||or (like instead of &&and). Interesting comment, thx!
So do manufacturers even make money at all from consumer grade procs? It's hard to see how they could considering everything. I always wondered why enterprise procs costed tens of thousands of dollars. Obviously they are better / less error prone / etc, but are they really a hundred times more expensive better? Now i'm guessing this is how they manage to make consumer grades affordable.
they do, otherwise they wouldn't exist... all you have to do is use google for 30 seconds and find out intel has a net income of 21 billion dollars in 2019
@r3d0c So what? That doesn't automatically mean that they make money on consumer grade products, and certainly doesn't preclude them from existence. The majority of that 21 billion dollar net income very probably comes from enterprise products (like servers, licensing agreements, etc). These types of things cost tens of thousands of dollars per unit. Inflating these prices would allow consumer grade processors to be manufactured and sold at a very slim margin.
They don't show us this in university, we have to learn 8086 architecture and memorize it... are you getting this? I have to memorize architecture of an ancient processor... I understand, I would appreciate if teachers just mention 8086 and even how it works but just briefly. It's hard for me to waste my time understanding it in depth as a SOFTWARE engineer...
Same :(
The predictions were very very accurate.
We need follow up for this
Wow, thanks for sharing
Next time you throw your old computer in the recycle bin, think of this video.
But that's just it, your old computer is literally rubbish. Its CPU is a square centimeter of sand. Its what it represents, or in this case what it did represent, which fast forward 10 years is superseded by magnitudes what a modern square centimeter of sand can accomplish. In material cost, even when it was created it was only worth 5 cents.
The material may be only 5 cents but the equipment and technology that was required to turn this small piece of sand into a working chip is worth billions of dollars, and years of work hours by some of the best minds humanity can offer.
5 cents for a wafer cut in how many pieces? 10.000? maybe 100k
@Cipi SixZeroFour Try 10's of billions... Maybe 100's if you're reading this from the future!
I'm waiting for bio technology , much more effective an simpler to use...
And yet nature produces computers with orders of magnitude more power for the price of a beer. Humbling.
Does anyone know where to get the slides?
http://www.toddfernandezinc.com/about/about-todd-fernandez/hope-number-nine
at the bottom of the page are the Google Slides.
Yeah... About that "Intel 7nm in 2017"...
@Minimal MO I never knew that, TIL, thanks
14nm+++++ in 2020
@toy machine But how does a fanboy make his way into a lecture about the details of semiconductor manufacturing? I would think this is probably way past your level, isn't it?
@Hjembrent Kent fanboy? that comment wasnt a fanboy comment. So the question is actually how get you here, if you dont even understand the simple concept of "fanboys"?
@Hjembrent Kent That wasn't a fanboy comment though? Just because someone makes a joke about a certain company doesn't mean they're fanboy's for the companies rivals, and frankly suggesting that makes you look like a fanboy.
Larry Gall - 2018-12-12
Even though it's from 2012, I found this one of the best videos on process that I've seen yet.