EUV has always been about achieving high enough power to be economically viable. It was never about making chips at any cost.
I remember reading the tinfoil hat theory about three-letter agencies making low-quantity high-cost chips at incredible process sizes in order to break encryption. I doubt that's still as viable today as it was before leakage currents started dominating, but it was an impressively plausible theory.
IIRC EUV development picked plasma over synchrotron because plasma projected to be cheaper, even though technically synchrotron had more benefits. Queue many, many years of solving for technical challenges for LPP and now commercialized EUV machines cost 200m, 400m for next high NA. Which is about the cost of multiple small or single medium size synchrotron facility. It's amazing plasma EUV works, but it's also a failure in the sense that it is FAR less economical than originally envisioned, which explains why particle accelerator route is still being worked on.
Back in the day, HP advertised that the distributed amplifiers in their 26.5 and 50 GHz equipment were made with e-beams, but the process size wasn't anything special, certainly not by today's standards. I'm not really sure what drove the decision.
that tinfoil hat theory, just as basically all of them, can only be produced by people that have absolutely zero understanding of the topic. The amount of challenges that industry has faced during the relatively fast progress through nodes is just non skippable, as there were so many things to be discovered through very expensive and long brute force (just one example: high k dielectrics)
No, you can absolutely make specialized chips that are orders of magnitude better than the commercial state of the art if you don't care about mass production or operational costs.
I can bet there are superconductor/photonics/topologically different/strange memory/smaller process size prototypes around.
Right now we are getting to the limits of transistor sizes, but even a couple of years ago experimental prototypes of smaller process size were developed years before mass production.
Well sure, at many orders of magnitude more cost. If you wanted to brute-force RSA, you'd be better off leveraging economies of scale to operate ten thousand current-gen 4nm GPUs running at 2 GHz with air coolers than a single exotic prototype 1nm ASIC running at 8 GHz with cryogenic cooling.
Aside from LCS35, most cryptographic problems are about as easy with two processors that are half as fast as one processor that costs twice as much.
Well I'm not an NSA engineer but I can imagine complex real-time analysis of streaming data, think processing 400+GbE link data. It quickly becomes too much too store, analysis could impose sequential packet dependencies.
proof? i'm just like... where? where do you think people are making these chips and using which ovens?
> smaller process size prototypes around
no there aren't. there just aren't. you could hide this in your basement about as easily as you could hide building your own space shuttle (and launching).
There's no hiding anything here. You can find random articles about publicity stunts companies used to try based on these. Specifically when IBM was in the game, and Intel until they got butchered by incompetence.
You can always turn a claim into a logically equivalent claim of the non-existence of any counterexamples.
“For every instance, e equals mc²”
is logically equivalent to
“There is no instance where e does not equal mc².”
That combined with your belief that claims of non-existence can't be held with any degree of certainty means you believe that no claim can ever be held with any degree of certainty. Which is not a very interesting insight.
This is an equally unsupportable claim, though. This requires enumeration of the entire state of the universe, an impossibility. This is just the standard swan problem (https://en.wikipedia.org/wiki/Black_swan_theory). What you have is a model you're very confident in without the deductively-rational basis your diction implies.
People should really read more hume if they're going to weigh in on philosophy of science.
> You do realize we all know it's impossible to have any degree of certainty in asserting the non-existence of something, right?
I love condescension that is so petty it's laughable. As the commenters said below there are very well-understood precedents/principles that allow me to conclude "no" here eg
So no it's not "impossible to have any degree of certainty in asserting the non-existence of something", we actually have a whole branch of mathematics dedicated to exactly that (it's called probability and statistics).
Probablility and statistics are models that produce something we call certainty. This has no relation to actual certainty, aka knowledge. If you're making an abductive claim, you should state it as an abductive claim. Otherwise you're simply claiming true knowledge that is literally impossible to have.
Too late to edit, but probability and statistics do emphatically rely on past-certainty. The entire concept of using the past to predict the future, however, is just a convenience with no reasonable basis. Please appropriately hedge your comments as to not imply otherwise or be appropriately mocked in response.
This is precisely why I don't trust people who aver without receipts to show. Open the schools, goddammit!
Until I see some reasonable evidence that smaller process size cannot exist, i just see lazy people getting angry that someone disagrees with them. All of this "burden of proof" bullshit, aping like you're in some kind of formal debate rather than a conversation with a stranger, just screams "emotional asshole who can't deal with someone disagreeing with them and never learned how to engage in basic conflict resolution when they had the ability to engage in good faith and chose not to".
Y'all deserve all the mockery society can afford. I'm at least honest in that I see conflict is what we need more than ever if only to put people like you in your place.
Hell, even liquid nitrogen temperatures are fine. More hassle than you'd want in your pocket but yearly running costs wouldn't be too bad for most businesses.
It's a superconducting switch, not a semiconductor per se. Lookup "Josephson junction". IBM spent a fortune on a huge R&D program to make computers from this, but eventually abandoned it. I think they got some circuitry working but eventually decided it wasn't practical enough to commercialize.
Also, some of today's work in quantum computers uses superconducting qubits. Maybe that's in the same research stage now. No idea if it will ever become practical.
Free Electron Lasers have potential to generate more tunable radiation with higher luminosity. Despite this they aren't a drop in replacement for the current EUV light sources. A free electron laser is 200 meters long, so a single laser would feed multiple EUV machines for it to be economical. This technology is very promising but it has been under development for a while. Does anyone know what the current difficulties are?
As far as I understand it, smaller scale XFEL devices still suffer from poor aim, even though now these machines have been miniaturized to basement scales. They don’t need to be significant fractions of a kilometer anymore. This aim issue will probably be solved in the next few years. It’s an exciting time to be in X ray science, particularly anything ultrafast.
Headline stuck me funny; I was working in ion implantation 20 years ago. Of course they’re talking about lithography, because those guys are the fighter pilots / first violinists of semiconductor manufacturing, they get all the attention.
Seriously! Although I’d say the most undervalued is wet processing. It’s incredible how much art and science go into not leaving water spots. Not to mention cleaning up other people’s messes.
No, we cannot. IBM moved neutral atoms around on an inert surface. No one has demonstrated building covalent structures (or metallic, or ionic for that matter).
My startup is trying to do this, and it is a fiendishly hard problem.
Is this "we" you and your startup or all of humanity? There are a variety of published papers that show simple memories and other structures that are way outside my domain knowledge (qd transistors?).
Why is it hard? You need to be able to position things with sub-angstrom precision from a platform that has ~nm uncertainty in the critical z positioning, and in the case of nc-AFM is oscillating to boot.
And you can’t use existing tools. You need an atomically precise scanning probe tip with very specific reactive chemical structure, but NOT react with the surface while scanning with a voltage bias.
And where do you source feedstock from? Needs to be delivered to the surface in passive form but be activated when needed to switch to being chemically reactive in a specific way to get it on the transfer tool and then onto the part being built.
Oh, and this is without even getting into how many electronic structures are entirely invisible at certain voltages, everything looks like an identical blobish shape, surfaces are reconfiguring themselves constantly, and probes randomly crash due to piezo creep, destroying days or weeks of work.
My startup has solutions to all of these problems. And the payoff at the end is reliable, scalable quantum computers, followed by full-on Drexlarian nanotech. But yeah, it’s a fiendishly hard problem.
The story of technological progress is one of shrinking feature sizes in manufacturing. Not just semiconductors, but everything. The Industrial Revolution is really the story of higher tolerance and more reliable manufacturing pins.
You can explore the physical limits of technology by looking at what happens when we reach perfect atomic precision--every atom where we want, in any configuration permitted by physical law. Across nearly every vertical, this represents a 100x to 1000x improvement. In some cases factors of 10^8 to 10^12 over present-day capabilities.
Developing a process to build structures atom-by-atom (essentially 3D printing diamond or other gemstone materials with atomic precision) would enable skipping to these theoretical limits, with the corresponding step function increase in functionality.
It would also move our technological base off being based on rare metals and alloys, and onto an industrial economy built on carbon (diamond and graphene), and other elements commonly available in the Earth's crust and atmosphere. After 3,000 years we will finally move from the Iron Age to the Diamond Age, and with it bring an eventual end to material scarcity and the economic basis for global conflict. You'd seriously need to go back as far as the invention of agriculture or Bronze Age or early Iron Age metallurgy to find a comparably transformative technological advancement.
Within the VC-fundable horizon of the next couple years, early versions of this manufacturing tech will permit making high-value quantum devices like sensors or qubits, as these can be manufactured by introducing certain defects into a growing crystal, with atomic precision relative to other defects or surface features.
What’s the plan for dealing with cosmic rays? I worry about when your beautiful angstrom-precision qubit networks encounter a relativistic proton or muon.
At near surface level ( 80m above ground in clear dry air ) 42 litres of doped Sodium Iodide scintillation crystal will experience ~ one to two thousand gamma events a second .. most of relatively low energy (and ground sourced).
The fall off from low orbit to surface is substantial in both event numbers and energy level.
The higher energy cosmic sourced events at surface level are down in the hundred or less a second (IIRC).
If there's a plan it'd likely include having 9x redundancy hardware surrounded by water deep in a former salt mine .. that'd take cosmic ray events way down and provide a (best of three) x (tell me three times) "just in case" statistical sharpening.
None of that is needed. You're talking about surface events per square meter (roughly) and we're talking about a device with total dimensions smaller than a single TSMC 2nm transistor. The cross section is so small that the chance of it being hit over the lifetime of the product is ignorable. There are way bigger operational risks to worry about.
At an irrelevant scale. The cross sectional area of these devices will be 18 - 20 orders of magnitude smaller.
> Always a possibility under consideration...
We're talking about the cross section of a macro-scale (visible with the naked eye) chip vs. a cluster of a few dozen atoms. Certainly you can understand the difference of scale? Cosmic ray induced bit flips are extremely infrequent events at the datacenter scale.
What's the frequency at which a single, specific transistor will be struck? Not that a bit flip occurs somewhere in a large datacenter, but the chance of just a specific transistor being hit. Now reduce that 100-fold. That's the base rate we're talking about.
The liklihood of a particular structure being hit is very, very small. Negligible over the operational lifetime of the device.
In the long-term vision of scaled-up nanotechnology, there will of course have to be redundancy and mechanisms for disabling, removing, and recycling (or incinerating) mechanisms destroyed by cosmic rays.
Things don't like to move once they're atomically-stuck together. Getting them to stick is another issue altogether. Doing so in reliable locations repeatedly at scale? Good luck.
Yes, electron-beam lithography is fantastic but also fantastically slow. Sorta like building a Lego model brick by brick vs layer by layer. It's still used for reticle fabrication and repair.
If you want to make features tinier than EUV allows, you do what you have done for the last few decades and make them directly, but real slow and costly, with electron beams. IMO at some point it seems likely that someone will simply decide to brute force e beam litho up to mass production rates.
Direct-write lithography has been a thing for a long time such as EBL. It's just SLOW. So it's only really viable for devices that are made in low quantities, simple devices or research.
I was hoping to see table top particle accelerators like those at UCLA were progressing into something usable for lithography. Which makes me wonder, why not use electrons instead of light?
From my non expert understanding, we already do kinda. The masks used for photolithography are made using an electron beam, allowing for a much greater resolution than what the underlying photolithography allows. But this is far too slow for large scale production.
Scanning an electron beam, repeatedly over an entire waffer would take forever. So instead we do it once, to make the mask, and that mask is then used over and over to expose the waffer.
This is a bit little injection molding: the mold is very expensive and made with a far better manufacturing process than the plastic pieces that it will eventually produce, but this is the price to pay for high volumes and low costs.
Adding to this, from what I’ve read electron beam is too slow for the required throughput. The ASML EUV machine can etch something like 170 wafers per hour. Using an electron beam would be far too slow for 2-3 wafers per minute.
It would be interesting to see if this tech was viable for dev boards. i.e. when you want to design a new 2nm chip, what if you were using electron beam chips to test out designs?
E-beam lithography is used for research purposes all the time, so yes it's already used for "dev-boards".
That said, apart from the economics (it is very slow), you are also constrained with respect to the size of what you can write, the writing fields is typically quite small, if you want to make a chip larger than that you need to stitch fields together and you have to deal with stitching errors (which becomes more and more difficult the smaller your structures are).
That would almost certainly be more expensive. When people talk about modern masks being $15M or so for each mask set, a huge fraction of the cost is this process for the masks.
It also doesn't tell you as much about how your design actually runs on the process in question.
That means each litho tool prints 6.75 x 10^13 “transistors” per hour. In more useful units, that’s 18.75B transistors per second.
Drawing them one line at a time is technically feasible but…I’ll bet you are talking single digit DIE per hour if that.
And that is for one layer of lithography. I’ve seen estimates from 5-20 layers of lithography using EUV tools at the 3-7nm nodes. So the time scale is even more warped.
It might be as simple as the fact that anything the electrons hit will pick up a huge electric charge. Now you've got ESD problems from hell, not to mention unwanted X-ray generation.
you can use an anti-charging layer for e-beam litho, that's not such a big deal. E-beam litho is just very slow. There are lithography techniques that use synchrotron sources like LIGA.
Hah, I think there’s two things that stand out in my memory.
- They need 3 Boeing 737’s to ship an EUV machine.
- We talked with one guy who’s responsibility it was to design one of the calibration points the machine uses to find it’s zero position. This left me amazed that they’re able to ship a machine halfway across the world, re-assemble it and calibrate it again to such accuracy. And! On top of that, make it reproducible over different machines!
Oh, I've actually seen that in videos! If you Google it, there's tons of pics showing the loading/unloading. They're not one-offs either. TSMC may order 80 at a time.
Hadn't thought about calibration afterwards. Crazy.
or some similar kind of device that turns the momentum of electrons into light. I'm a little surprised that they didn't try something like a FEL first instead of that terribly problematic device that uses highly inefficient lasers to blow up tin droplets, itself a high-loss process that produces contamination and resulted in years of delay developing materials for
They tried both in the initial design phase, there's upsides and downsides, but ultimately thought that the tin droplet laser was more liekly to actually get done more or less on the time schedule requested, and so that's where the bulk of the capital went.
Interestingly China has been continuing working on the synchotron based EUV litho idea (in addition to work to create domestically built tin laser EUV lithos machines).
My bet is on plasma Wakefield accelerators to feed the FEL. But yeah a synchrotron might do as an intermediate step. Free Electron Lasers can be tuned to different wavelengths all the way to x-rays.
I remember reading the tinfoil hat theory about three-letter agencies making low-quantity high-cost chips at incredible process sizes in order to break encryption. I doubt that's still as viable today as it was before leakage currents started dominating, but it was an impressively plausible theory.