Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

> For instance, compared to an EUV system employing a 250-watt light source, Canon estimates NIL consumes just one-tenth the energy.

I'm not an expert on this but feel like a 250w light is not the major driver of cost in EUV? Or am I misunderstanding this?



Producing 250w of EUV light requires 20+ kilowatts of electricity pumped through an extremely expensive system of lasers and mirrors.


AFAIK the main coat of EUV is cost of machines that will be obsolete in few years, so you want to produce as many chips using them as possible during that timeframe, they design a lot of around to maintain near 100% uptime of those machines. This include buffers before and after machines (so any unplanned stalls are mitigated) and technicians trained to do maintenance in F1 pitstop fashion. (source: some tour of some chipmaker I saw online, no longer remember details)


> to maintain near 100% uptime

They have uptime only about 80%. They need to be stopped, calibrated and maintained frequently.

They do not go obsolete quicly. They are constantly upgraded. 10-15 year old fabs and machines are still running all over the world. There are 1000 nm, 90nm, 40 nm, 14 nm fabs still running. High-end is not all of semiconductor industry.


Yes, it is more that margins are highest when a machine is brand new so it pays to maximize duty cycle. It will last for a decade, sure, but those 90nm fabs are not nearly as profitable as a 4nm fab.


That's not because they become obsolete, it's because they're the rate-limiting step (bottleneck).


These machines will be not obsolete for very long time. They are extremely rare and expensive. And the most of semiconductors are fabricated on mature nodes anyway.


Not obsolete, but the margin on fabs drops off pretty quickly once they're not at the cutting edge, as I understand it, so they need to make back their capital investment fairly quickly.


With exceptions, was reading recently an interesting die analysis/estimate of the costs and margins of manufacturing of the AD9361 chip (a 65nm, digital radio transceiver, introduced 12 years ago and still selling at retail today for $300/$400):

https://zeptobars.com/en/read/AD9361-SDR-Analog-Devices-DAC-...

Relevant quotes (and the current retail price if anything is higher now then when the article was written):

“ Retail price of AD9361 at distributes is 275$, volume price from manufacturer is 175$.

That is quite an impressive added value! For 1,68$ of manufacturing cost we are getting 173,32$ of added value! Even Intel with their x86 or drug cartels could NOT do business like that.”

Of course, the actual margin needs to take into account NRE and other costs (and the above link does get into that) but, in this case, the manufacturing is a tiny sliver of the costs.


When an ASML Lithography Machine Goes Down: https://youtu.be/6v9gx3Z4oVk?


"Obsolete" which I guess for you means for the bleeding edge? Larger nanometer processes will still be in use since their cost will come down. For example when automakers stopped their orders for chips during COVID they pivoted (ported?) to higher nanometer designs because it wasn't a core requirement.


AFAIK, 250W is the net energy of light arriving at the wafer after it has reflected off of many mirrors, with a very inefficient process to generate light from the tin plasma on top of that.


Yes that's a strange quote. 250w is not only wrong, but absolutely minuscule compared to the tech around it. I work on the optics system


Weird article. The energy consumption of an EUV machine is about 1MW, that's why it's interesting to have an efficient alternative, not the actual useful power of the source.


That's why they're working on FEL light sources.


That sounds like a lot, but how many of those machines does an average fab have? Seriously I have no idea.


I have visited 3 fabs, only one machine in each. Anecdotal info…


In that case 1MW doesn't sound like a whole lot to be worth optimizing for.


According to Claude, that's about $1M/yr of electricity, assuming 24x7 usage.

I assume the real saving is on the cost of the machine in the first place, and again relying on my AI buddy Claude:

Let me break down the costs of both Nanoimprint Lithography (NIL) and Extreme Ultraviolet (EUV) lithography machines:

NIL Machine Cost:

Basic NIL systems: $1-3 million Advanced NIL systems (like those from Canon/Molecular Imprints): $10-15 million

EUV Machine Cost:

Current ASML EUV systems (like the NXE:3400C): Approximately $150-200 million per unit Latest generation ASML EUV systems (NXE:3600D): Over $300 million per unit Installation and support infrastructure can add $30-50 million

**

So, looks like $200M+ saving going with NIL vs EUV.


The driving CO2 amplifier should be already beyond that figure alone


Yes, it is 250w output but the efficiency is near zero, really


But you can bake a cake with a mere 60w light. That 250w is also much less than what it takes to run modern GPUs. so it's all relative.


The EUV light is produced by shooting a pulsed laser on tin droplets.

You already lose most of the input power in the pulsed laser. Then only a fraction of the energy of the light hitting the tin is converted to EUV light with the correct wavelength. Finally the EUV light has to be focused on the mask through complicated optics, which is notoriously difficult for EUV light.

I guess, there are other sources of inefficiencies, that I forgot.


I’d just like to comment on how batshit insane the technology is.

“We pulse lasers in sync with dispensing droplets of molten tin to produce light that doesn’t exist outside of stars, then we use mirrors with a sub-angstrom surface roughness to precisely direct it onto wafers.”

Not to mention the fact that this is happening, IIRC, thousands of times per second, and the tool has to take the wafer’s topography into account to focus the beam. Honestly, EUV litho makes every other technology you could describe sound like child’s play.


Roughly similar craziness: Disk drives mechanically position heads less than 1nm above the platter, with horizontal accuracy of significantly less than the 50nm track width, at a retail price of a few hundred $$ or less.

30 years ago I think you could have gotten any number of experts to explain why both EUV lithography and modern disk drives are impossible.


True, and all while the platter is spinning at thousands of RPM, perfectly balanced. HDDs are wild.


And the tin droplets are in vacuum, with contamination of the wafer being very critical to control.



The first time I read about this process, I was convinced aliens were involved. Seriously, it's one of those crazy pitch meeting things that sounds ridiculous so of course it was green lit. "So we fire this laser, pew pew, into a field of molten tin droplets, and bingbangflam, you get this flash of light. So what do you think?" Hold my beer.

It's clearly some people that are very smart that can only be explained by aliens


Is there a reason EUV labs don't collocate with synchrotron FEL lasers?


Indeed, synchrotron sources are considered to be a possible light source for lithography. China is especially interested in them, since they can be done with a homegrown technology: https://www.eenewseurope.com/en/chinas-synchrotron-euv-litho...

From what I understand, tin-based sources are easier to work with because they are point-like. All the energy is produced from a tiny droplet. Synchrotron sources produce much wider beams, that need to be re-focused properly.


Synchrotron FELs are already used, and their construction paid for due to their scientific uses. Building new ones would require new synchrotrons, which are more expensive than the crazy "hit molten tin drops with lasers to make a plasma" scheme the ASML machines use.


Yes. Nucleophobia.

Same reason it's called "euv" and not "soft x-ray".


"because the bit flips would then already have infected your processor in the factory" ?


Isn't it actually tin plasma?




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: