Inside the miracle of modern chip manufacturing | Hacker News Reader

Inside the miracle of modern chip manufacturing

ig.ft.com

125 pointspseudolus23 comments

I was curious that they claim that lithography is done with ultraviolet, and I looked up the wavelengths involved. Sure enough, "typical" UV is still hundreds of nanometers long, like visible light. This would be too coarse for small features since you'd get diffraction at the edges of the stencil. I looked up UV in wikipedia and apparently they have defined "extreme UV" to be down to 10nm - which I would have called X-rays. Indeed, if 10nm is extreme UV, perhaps we can also call 10nm light "underachieving X-rays".

https://en.wikipedia.org/wiki/Ultraviolet

Ok, so the answer is marketing, but there's a long story. There were so many failures of X-ray lithography that it was easier to name it EUV. Sorta like you don't have nuclear MRI for medicine. They choose to emphasize the 3D Imaging over the n-word.

One can get quite small even with DUV (193nm ArF) by using high NA (water immersion) to get 1.3-1.4x better resolution (wavelength scales with index of refraction). That along with multi-patterning (which puts limits on the design layout, but allows 2-4x tighter pitch) can get you down to a pitch of about 65nm (single patterned) or 10-20nm (multi-patterned). However, the whole idea of pitch for the gate length falls apart around here and 40 is basically the same pitch as 65 with transistor packing tweaks, and 28 is the last "analog" node. By the time you're multi-patterned bellow 22nm, you're using fin-FETs and the scaling is totally broken (some would say made-up by marketing). That's how TSMC's 7nm was the same pitch as Intel's 10nm. Quad multi-patterning's poor yield finally broke the long dominance of 193nm, but it was cheaper than X-rays for about 20 years (and a lot of bankruptcies).

The use of the name EUV with 13nm wavelength and now immersion EUV, which allows 2-3nm without multi-patterning is meant to evoke the extension of lithographic techniques rather than the revolution X-rays implied 20 years earlier. Note that it's still so expensive that only a few critical layers are patterned with EUV. Double patterning will double the costs (half as many layers patterned per stepper per hour) so all the old tricks will come out again. Transistors will evolve again with nano-sheets and GAA to reduce leakage at even closer spacings. We should get below "2nm", but you're not talking about gate or metal dimensions anymore, just peak transistor density is ~400x better than 40nm (1M/mm2 vs 400M/mm2).

When you're selling $100M pieces of equipment in a field wit decades of history, names matter.

An intriguing answer, but I'm afraid I didn't know enough terms of art to really understand it. Here are some definitions for the interested:

ArF: argon fluoride. ArF lasers emit at 193nm

EUV (Extreme Ultraviolet Lithography): A type of lithography that uses light with a wavelength of 13nm

DUV (Deep Ultraviolet Lithography): Refers to lithography techniques that use deep ultraviolet light, such as 193nm ArF, for patterning.

NA (Numerical Aperture): A measure of the light-gathering ability of an optical system, like a lens, used in lithography to improve resolution.

Water Immersion: A technique where water is used between the lens and the wafer to enhance resolution in immersion lithography.

Multi-patterning: A method where multiple exposures are used to create smaller features. (Would like to know more)

Pitch: The distance between identical features in an array, crucial for determining the density of components on a chip.

Gate Length: The length of the gate in a transistor - affects its performance and power consumption.

Fin-FETs (Fin Field-Effect Transistors): Transistors with a fin-like structure that improve control over current flow, commonly used in advanced nodes.

Nano-sheets: a three-dimensional transistor structure that utilizes thin horizontal sheets of semiconductor material stacked vertically.

GAA means: "gate all around" - the gate surrounds the channel on all four sides with stacked horizontal nano-sheets.

> Multi-patterning: A method where multiple exposures are used to create smaller features. (Would like to know more)

It might be better to start by considering ways to play with shadows cast by a single photolithography mask...

The Wikipedia article on "Computational Lithography" is a good place to start:

https://en.m.wikipedia.org/wiki/Computational_lithography

A particularly fascinating technique, linked in that article, is the use of "Phase-Shifting Photo Masks":

https://en.m.wikipedia.org/wiki/Phase-shift_mask

And then there are ways to play with diffraction effects, at the edges of the opaque shapes of the mask:

https://en.wikipedia.org/wiki/Optical_proximity_correction

In addition to diffraction and phase shift patterns, you can create shapes by illuminating the mask and wafer with more than one light source.

It may be helpful to consider this "multiple illumination" of a single photo mask in the context of a "double slit" interference pattern:

https://en.m.wikipedia.org/wiki/Diffraction_from_slits#Multi...

After you've run out of tricks, you have write over the shapes laid down by a single mask, running the wafer through the process with a different mask.

I find the current Wikipedia article for multi-patterning to be a bit more difficult to understand, but it's certainly worth the effort.

After studying this article a bit, I think it might help to start in the middle of the article, at "Pitch Splitting". A second read of the article, starting at the top, may then provide more context:

https://en.wikipedia.org/wiki/Multiple_patterning#Pitch_spli...

(Ironic: the article on Multiple Patterning requires multiple passes!)

fantastic info. thanks for sharing.

Hah, I left actual silicon process engineering a long time ago, but I worked on the early 68020 at 2um and then early submicron in R&D. That was 40 years ago and the process now 1000x smaller (1 million more transistors). It's pretty amazing... even if it does seem like fab investment costs have grown so much that we're nearing the end of profitable scaling.

> I was curious that they claim that lithography is done with ultraviolet, and I looked up the wavelengths involved. Sure enough, "typical" UV is still hundreds of nanometers long, like visible light. This would be too coarse for small features since you'd get diffraction at the edges of the stencil.

173nm DUV is used commercially to create features in the 20-30nm range, possibly smaller.

> I looked up UV in wikipedia and apparently they have defined "extreme UV" to be down to 10nm - which I would have called X-rays. Indeed, if 10nm is extreme UV, perhaps we can term 100nm light "underachieving X-rays".

What are you are trying to say? EUV light used in lithography (13.5nm) is close to soft x-ray light, but what does 100nm light have to do with it?

It was a typo, now corrected. I was suggesting, humorously, that 10nm light could be called both 'extreme uv' and 'underachieving xrays', since they overlap. The EM spectrum is continuous, so these distinctions and names are arbitrary, so why not have some fun?

Well, I would have called infrared microwaves because the wavelengths are in micrometers, but I believe that ship sailed over a century ago. (P.S. they're using them to cause chemical reactions, not ionize atoms or scatter off electrons, which is definitely more of a UV thing than an X-ray thing.)

> they're using them to cause chemical reactions, not ionize atoms or scatter off electrons

Those three are approximately the same thing.

The distinguishing characteristic of X-rays is that they remove electrons from other layers than the last one. And the most relevant feature of UV is that it ionizes atoms. (Even though what distinguishes it is that we can't see it.)

For completeness, the thing that differentiates IR from microwaves is that IR creates molecular compression while microwaves can only make matter vibrate.

The IR name is about 3 centuries old, and it's because it often comes from light sources. While microwaves are the smaller version of the short waves; from the short, medium, and long waves used on radio.

Anyway, the dividing between those is always fuzzy.

Would anyone know the general steps to making an interactive webpage like this? I tried to take a peek by opening the Web Inspector and scrolling to see what changed, but couldn't uncover anything definitive.

The general technique I can identify is that the webpage uses an approach called "scrollytelling" (also noted by the references to "scrolly" from the Web Inspector) to make the animation advance in reaction to the user scrolling down.

For the pictured objects, I assume that an animation team (likely a bit separate from the web development team) created the assets. But I wonder how the developers then made sure that each provided asset enlarged at the proper rate as the user scrolled down (and also made sure that the scaling worked properly upon changing the width of the window).

https://pudding.cool/process/how-to-implement-scrollytelling...

We're not scrubbing over a pre-rendered animation here. We're using a library called threejs (a WebGL wrapper) in the background figure and adjusting the camera/models in response to scroll position (within a react environment)

That is excellent, thanks for the information and work to produce a very nice feature.

I still remember a time when radios occupied almost a whole table, and the magic of portable "transistor radios". we used to smuggle them into class to listen to the broadcast of cricket matchew. The teacher would pretend not to know we were listening to it, but when a wicket fell, so did the facade, and he'd ask what the score was. we would tell him, the whole class unified in a conspiratorial code of silence

When viewed from distance, progress we had in last 100 years is just magical, and sometimes there are local pockets of old technology. My father told me story about how he learned electronics from an old book where they started with a "crystal radio" from scratch (like finding a conducting region in some mica crystal) and he tried to buy some old diodes in local electronics shop and they were already obsolete and they had much better diodes already on stock. When I started, I had to buy some discrete transistors and amazed him with minimal quantity order of 20pcs, and I bought 100 just because it was still pennies. Recently I was amazed by ssd's being so cheap. We have A LOT of progress, but we have literal explosions of progress in adjacent areas that diminish that little slower progress, so people say "oh, yeah, the progress times are behind us now".

I suppose the '47' in your username tells the same story as the Kalashnikov one does?

sorry, I am not familiar with the Kalashnikov story, so it would be an incredible coincidence if it did!

Good storytelling, but lacking editorially - they left out an important category of chip designers / manufactures and missed out on perspectives from companies that design chips as part of their products but not as end products. Leading to flawed assertions that (all) engineers are moving away from System on a Chip(SoC) architecture to chiplet architecture & multi-chip modules (MCMs) technology! From the likes of intel, it makes sense to modularize their product and maximize market outreach by offering variants. But this would almost always be subpar efficient on power and performance parameters to SoC designed for custom use-cases!

Ironically even this article is a decade behind as it depicts planar CMOS transistors. Around 2010 the industry started moving to finfets which are gates surrounded on three sides by poly. Manufacturing those structures almost destroyed Intel as the first process had their worst yields in history. Source: former Intel process engineer that got out before p1274/1276 shat the bed.

From a customer perspective, though, Intel's first FinFET chips (22nm, Ivy Bridge) were at most a few months late, and price and performance were reasonable. The next node was a bit more late and definitely underwhelming at first, and the one after that was the real disaster.

I wish there would be a flattened alternative in order to read the text without distractions.

It's called "a book". ;-)

Reply on news.ycombinator.com