It's been mentioned before, but Chris Miller's Chip War from a few years back is an excellent, very-readable book on the topic. Goes into depth on the history and development of chips and their production. He did the rounds on the interviews back then, and it's definitely worth a read. The EUV stuff is great, but I particularly liked his history on how the USSR was always going to lose and how integral Apollo really was.
Yeah… when I’m eating breakfast, a lecture is not what I’m after. I watched that Veritasium video a while back and was glued to it. Any other presentation style and I probably would have completely skipped it (thinking I’ll watch it another time knowing I would never go back to it).
> These machines are roughly the size of double-decker buses. To ship one requires 40 freight containers, three cargo planes, and 20 trucks. They are the world’s most complex objects. Each contains over one hundred thousand components, all of which have to be perfectly calibrated for the machine to produce light consistently at the right wavelength.
As a software engineer by trade, the above parable communicates to me two very important things and little else by comparison: that the machines are ultimately fragile and nowhere near "optimised", since the complexity is by own admission substantial to put it mildly; the machine is not a commodity, exactly, one of the million pieces breaking subtly likely renders it inoperable; its cost is proportional to its complexity (read: astronomic); by mere fact it's a focal point of geopolitics only supports the rest of the argument it's a machine of current stone age much like siege engines were at some point the closely guarded secret win-or-lose multiplers of feudal culture.
I mean it's certainly interesting to read about the complexity, but reducing the complexity and commoditising the whole thing is what's really going to be impressive I think :-)
I am probably speaking out against the nerd in us, and none of what I said should detract from enjoying the article or the subject, it's just that I think complexity here is the giveaway of us not having conquered UVL exactly, not quite yet :-) Or maybe we lack the right materials which would allow us to reduce the machine or make it less complex or prone to calibration related errors.
Indeed, all this reminds me of the marvel that is mechanical timekeeping - incredibly complex engineering that would ultimately be surpassed by dirt cheap electronics.
What is the corresponding revolution in chip production? I imagine something like FPGAs for litography - a wafer that can somehow work on another wafer in a sandwich-like configuration. Such a process could potentially improve on each iteration and thus get very good, very fast.
> By betting on extreme ultraviolet lithography long before it worked, ASML became the chokepoint for cutting-edge chips.
Makes one wonder: Would we be much better off of worse off if we reshaped society to do more of things, where a new technology is unlikely to work but highly beneficial in the limits? Would we sooner have 10 additional ASMLs or waste a lot of resources?
It is unavoidable that, at some point, China will have its own matching or better machine because they obviously how incredibly strategically important it is.
Non-zero chances - yes. Unavoidable - I wouldn't be so sure. I can't imagine how many top human-hours and cutting-edge inventions involved to construct this machine. And much of this simply cannot be stolen or bought, no matter how much money you have.
It has never happened in the history of the world that a company or country could maintain its technological advance indefinitely.
Either China will catch up on this or that particular technology will become obsolete. But it is certain that they won't stay behind forever (measured in a small number of decades at most).
i find it hard to believe that there is no equivalent anywhere else in the world. there is so much talent out there and the stakes are so high that it seems like an inevitability.
whatever many secrets are involved, information wants to be free and it's hard to believe that others won't figure it out.
by the time they do catch up we better be steps ahead. what's after EUV?
Honestly I thought the same, but after watching a couple of videos on how EUV actually works, and what ASML (and the 1,200 other specialized companies that feed into its supply chain) built..
I can understand why you can't just take one apart and copy it.
There's (apparently) 4 decades of accumulated cutting edge scientific research that has gone into these machines.
I suspect the machinery, process and human expertise required to simply produce the parts required for these machines is the real moat (oh and I guess the US-led export controls too).
The build tolerances for components are incredible. There are 11 primary mirrors in an EUV machine, each one has something like 100 coats of ultra-pure materials that are precisely deposited in picometer-thick layers with tolerances in the nanometers, across a 1-meter wide curved surface.
Then you have to position the mirrors perfectly inside the machine, again with tolerances in the nanometers.
So even if you know what you need to do, having the equipment and expertise to do it is a different thing.
And that's just one part of the 100,000+ parts that make up an EUV machine.
If there's really such a bottleneck around ASML, why not design some extra chips for legacy processes that presumably already have well known design workflows?
I mean we're not talking AMD FX and Core 2 Duo here, it's Raptor Lake and Zen 3, it's perfectly viable and still being sold in droves right now.
That’s what the likes of AMD with their chiplet design have been doing.
There’s also the issue of older process nodes not being profitable enough anymore, which explaines why at the height of the chip supply crunch older ARM chips were in short supply but there was ample stock of the 20nm feature-sized RP2040.
This is gonna sound super dumb, but I'm not sure how they aren't being profitable if there are shortages, just price things beyond break even level? The average person can't even tell the difference between a Core 5 and a Core 5 Ultra, you can practically sell them at the same price and I'm not even sure they'd notice when actually using them. The performance jump is relatively minor and the bottlenecks are elsewhere.
It mostly comes down to the consumer market not being significant enough by itself. A consumer may not notice a 10% increase in performance per watt or dollar. A large office building probably will, and a datacenter definitely will.
I don't think I'm being entirely hyperbolic when I say the consumer market only exists to put devices that can connect to and feed the datacenter loads into the general populations hands.
Isn't exactly this what China is doing? Apart from poaching ex ASML employees? Now reaching 7nm, and just throwing up more energy to catch up in FLOPS like Jensen said?
Because very large share of market now are datacenters. Difference from desktop is dramatic - for desktop really acceptable very simple chips with bad energy efficiency, but DCs already deal with extremely high power consumption, as they typically "compress" so much consumption in one rack, that constantly working near to physical constraints.
That's the AI hype narrative, but aren't server CPUs only like 25% of the total market? That's tiny compared to consumer volume, though revenue is likely on par given the higher cost per unit.
You can't make desktop computer 4 times larger but there's very little preventing you form putting 4 racks where you had 1 before. If the floor space is the expensive part of data center then probably some incentives are misaligned.
You cannot place dc anywhere, in large cities space is extremely constrained, and land is extremely expensive.
Also big problem - connectivity - you cannot place DC where it cannot be connected to power grid and to very powerful network.
So yes, DC floor space is severely limited.
And the third issue - last decades, rack servers dissipate extremely large amounts of heat, I hear numbers up to tens Kilowatts per rack, which is just hard to dissipate with air cooling (as example, all IBM Power servers have option of liquid cooling, but this is totally different price range).
AMD hiding Threadripper behind their back: Uh yeah what a terrible idea, we definitely didn't actually do that. Making a CPU that's twice the size, how ridiculous would that be right?!
It looks complicated but I suspect that 90% of what I see in that picture is just a giant refrigerator.
It's been mentioned before, but Chris Miller's Chip War from a few years back is an excellent, very-readable book on the topic. Goes into depth on the history and development of chips and their production. He did the rounds on the interviews back then, and it's definitely worth a read. The EUV stuff is great, but I particularly liked his history on how the USSR was always going to lose and how integral Apollo really was.
For anyone interested in the topic I highly recommend this Veritasium video: https://www.youtube.com/watch?v=MiUHjLxm3V0
The Veritasium video is good but their "newscast" style with constant back-and-forth cuts to talking heads can make the presentation a bit disjointed.
The more straightforward video of ASML EUV is from Branch Education: https://www.youtube.com/watch?v=B2482h_TNwg
Because that vid gives an overview of the whole machine, it gives context to what each scientist is talking about in the Veritasium interviews.
Thank you! The video you recommended definitely goes more in depth. I still like Veritasium's style more but it's just personal preference ofc
Yeah… when I’m eating breakfast, a lecture is not what I’m after. I watched that Veritasium video a while back and was glued to it. Any other presentation style and I probably would have completely skipped it (thinking I’ll watch it another time knowing I would never go back to it).
Great video and I think the only way to truly grasp the complexity of EUV lithography as a layman.
> These machines are roughly the size of double-decker buses. To ship one requires 40 freight containers, three cargo planes, and 20 trucks. They are the world’s most complex objects. Each contains over one hundred thousand components, all of which have to be perfectly calibrated for the machine to produce light consistently at the right wavelength.
As a software engineer by trade, the above parable communicates to me two very important things and little else by comparison: that the machines are ultimately fragile and nowhere near "optimised", since the complexity is by own admission substantial to put it mildly; the machine is not a commodity, exactly, one of the million pieces breaking subtly likely renders it inoperable; its cost is proportional to its complexity (read: astronomic); by mere fact it's a focal point of geopolitics only supports the rest of the argument it's a machine of current stone age much like siege engines were at some point the closely guarded secret win-or-lose multiplers of feudal culture.
I mean it's certainly interesting to read about the complexity, but reducing the complexity and commoditising the whole thing is what's really going to be impressive I think :-)
I am probably speaking out against the nerd in us, and none of what I said should detract from enjoying the article or the subject, it's just that I think complexity here is the giveaway of us not having conquered UVL exactly, not quite yet :-) Or maybe we lack the right materials which would allow us to reduce the machine or make it less complex or prone to calibration related errors.
Indeed, all this reminds me of the marvel that is mechanical timekeeping - incredibly complex engineering that would ultimately be surpassed by dirt cheap electronics.
What is the corresponding revolution in chip production? I imagine something like FPGAs for litography - a wafer that can somehow work on another wafer in a sandwich-like configuration. Such a process could potentially improve on each iteration and thus get very good, very fast.
> By betting on extreme ultraviolet lithography long before it worked, ASML became the chokepoint for cutting-edge chips.
Makes one wonder: Would we be much better off of worse off if we reshaped society to do more of things, where a new technology is unlikely to work but highly beneficial in the limits? Would we sooner have 10 additional ASMLs or waste a lot of resources?
I mean, it’s been tried; a reading of relevant historical texts would give you lots of ammunition to support either argument.
I'm pretty sure that's the most bizarre light source on the planet: https://youtu.be/B2482h_TNwg?t=929
It is unavoidable that, at some point, China will have its own matching or better machine because they obviously how incredibly strategically important it is.
Non-zero chances - yes. Unavoidable - I wouldn't be so sure. I can't imagine how many top human-hours and cutting-edge inventions involved to construct this machine. And much of this simply cannot be stolen or bought, no matter how much money you have.
It has never happened in the history of the world that a company or country could maintain its technological advance indefinitely.
Either China will catch up on this or that particular technology will become obsolete. But it is certain that they won't stay behind forever (measured in a small number of decades at most).
Right but if you dont say how long it will take them, youre not really saying anything.
i find it hard to believe that there is no equivalent anywhere else in the world. there is so much talent out there and the stakes are so high that it seems like an inevitability.
whatever many secrets are involved, information wants to be free and it's hard to believe that others won't figure it out.
by the time they do catch up we better be steps ahead. what's after EUV?
[delayed]
Honestly I thought the same, but after watching a couple of videos on how EUV actually works, and what ASML (and the 1,200 other specialized companies that feed into its supply chain) built..
I can understand why you can't just take one apart and copy it.
There's (apparently) 4 decades of accumulated cutting edge scientific research that has gone into these machines.
I suspect the machinery, process and human expertise required to simply produce the parts required for these machines is the real moat (oh and I guess the US-led export controls too).
The build tolerances for components are incredible. There are 11 primary mirrors in an EUV machine, each one has something like 100 coats of ultra-pure materials that are precisely deposited in picometer-thick layers with tolerances in the nanometers, across a 1-meter wide curved surface.
Then you have to position the mirrors perfectly inside the machine, again with tolerances in the nanometers.
So even if you know what you need to do, having the equipment and expertise to do it is a different thing.
And that's just one part of the 100,000+ parts that make up an EUV machine.
Maybe copying it is too fragile (but I note that China copied the F-35)
But in this case the Chinese will just develop their own alternative, that might work as good or even better
"at some point" is doing a lot of work there. How long do you think?
30 years
They might be the most complex mass-produced commercial machines but the Large Hadron Collider has a plausible claim to the title of "world's most complex machine" https://www.guinnessworldrecords.com/world-records/103591-la...
I think something like this wins that category:
https://en.wikipedia.org/wiki/Eastern_Interconnection
It's strange to count this as single machine. But if we go this way, the north Chinese interconnection and the continental Europe interconnection are bigger (https://en.wikipedia.org/wiki/Wide_area_synchronous_grid).
I suppose it's partly a semantic question that hinges on what you count as a single "machine" and what's a system or a network.
agree
and yet not even close to the complexity of the human brain
If there's really such a bottleneck around ASML, why not design some extra chips for legacy processes that presumably already have well known design workflows?
I mean we're not talking AMD FX and Core 2 Duo here, it's Raptor Lake and Zen 3, it's perfectly viable and still being sold in droves right now.
That’s what the likes of AMD with their chiplet design have been doing.
There’s also the issue of older process nodes not being profitable enough anymore, which explaines why at the height of the chip supply crunch older ARM chips were in short supply but there was ample stock of the 20nm feature-sized RP2040.
This is gonna sound super dumb, but I'm not sure how they aren't being profitable if there are shortages, just price things beyond break even level? The average person can't even tell the difference between a Core 5 and a Core 5 Ultra, you can practically sell them at the same price and I'm not even sure they'd notice when actually using them. The performance jump is relatively minor and the bottlenecks are elsewhere.
It mostly comes down to the consumer market not being significant enough by itself. A consumer may not notice a 10% increase in performance per watt or dollar. A large office building probably will, and a datacenter definitely will.
I don't think I'm being entirely hyperbolic when I say the consumer market only exists to put devices that can connect to and feed the datacenter loads into the general populations hands.
Isn't exactly this what China is doing? Apart from poaching ex ASML employees? Now reaching 7nm, and just throwing up more energy to catch up in FLOPS like Jensen said?
Because very large share of market now are datacenters. Difference from desktop is dramatic - for desktop really acceptable very simple chips with bad energy efficiency, but DCs already deal with extremely high power consumption, as they typically "compress" so much consumption in one rack, that constantly working near to physical constraints.
That's the AI hype narrative, but aren't server CPUs only like 25% of the total market? That's tiny compared to consumer volume, though revenue is likely on par given the higher cost per unit.
You can't make desktop computer 4 times larger but there's very little preventing you form putting 4 racks where you had 1 before. If the floor space is the expensive part of data center then probably some incentives are misaligned.
You cannot place dc anywhere, in large cities space is extremely constrained, and land is extremely expensive.
Also big problem - connectivity - you cannot place DC where it cannot be connected to power grid and to very powerful network.
So yes, DC floor space is severely limited.
And the third issue - last decades, rack servers dissipate extremely large amounts of heat, I hear numbers up to tens Kilowatts per rack, which is just hard to dissipate with air cooling (as example, all IBM Power servers have option of liquid cooling, but this is totally different price range).
Bigger chips = more distance to cover for your electrons = more power required = more generated heat = slower throughput for your data.
Surely you don't believe that the entire chip industry had not thought of "wait what if we just make the chips bigger".
AMD hiding Threadripper behind their back: Uh yeah what a terrible idea, we definitely didn't actually do that. Making a CPU that's twice the size, how ridiculous would that be right?!