I didn't understand how crucial chip manufacturing is until reading the book "Chip War".. What an amazing book. This is probably one of the greatest, if not the greatest invention in the human history.
Just the sheer number of companies and the complexity of the supply chain to produce a square inch of product most of which is tamed sand… is mind boggling.
I don’t think there is quite another product made by humankind that matches the sheer number of humans/sqinch - even including rockets, space ships, and some hospital clean room equipment.
Truly a golden era to be alive as a civilization to appreciate the height of human potential.
Almost every current era is golden if you lived in that era, unless you look back. Tech is always improving every year, if it isn't tech, is laws and social structures.
This is not true. The idea of constant progress is a relatively recent cultural invention.
Prior to the great acceleration changes were slow enough that the world seemed static over a lifetime. If it wasn't, the changes were almost as likely to be negative as positive.
It's also entirely possible that we are in a transitional period. That in hindsight the great acceleration will be the first half of a sigmoid function.
Makes me think of something my high school physics teacher tried to drill into us: if you zoom in on an exponent enough, it looks like a straight line.
Maybe we're transitioning between stages of a sigmoid. Or maybe we're still on an exponential trajectory, but on a timescale of one lifespan, the change still looks linear (if very fast).
When I was a kid I was into electronics and remember being very annoyed when I found out about this chip stuff. I was happily building analog circuits using the basic building blocks like transistors, capacitors and resistors and it gave me the satisfaction of having great control over everything and being able to "see in 3D" when I look at my circuit as I had full mental model over what's happening.
Chips are really great of course but at the same time its completely magic in a bad way. It haven't clicked for me how it all works until I watched videos of people building CPU analogs in Minecraft or something. Just don't like the feeling of not having an idea of how this thing works just by looking at it.
> Just don't like the feeling of not having an idea of how this thing works just by looking at it.
I find it pretty funny that, since we're no equipped to sens electrical flows just by looking at it, it's more the visual / sensorial support for you mental model that helps you. E.g. you don't really have more information by seeing the circuit physically than looking at plans, it just that your brain finds it easier.
But it's also pretty clear that having physical objects in the loop actually changes the brain thought process, so it's not just an affair of having the information at hand, how it's presented to our senses must imply changes.
That’s right, the analog circuit has clear topology which I can easily imagine what’s going on just by looking at it and sometimes by manipulating it.
It’s simply more enjoyable to look at, like looking at mechanical watches or industrial machinery. Chips are kind of dull. Not to take anything away from their enormous utility and impact of course.
Well, this us true for only the simple analog circuits. A modern analog front-end of a 5G modem or so is absolutely impossible to understand visually. Even the sub-blocks like the phase-locked-loop (PLL) or the analog-to-digital converter (ADC) are often vay too complex to grasp alone. But with a good hierarchy of the schematic one can divide each of these into sub-blocks and understand each of the sub-blocks. The whole is still too complex to understand fully, even for its designer. This causes a lot of simpe mistakes to happen like forgetting to connrct two nodes and not realizing till the silicon failing after the production...
Yes, but there is a certain magic feeling of power in knowing that you could, with a steady hand and a soldering iron with thin enough tip, meaningfully alter or repurpose the analog parts of the circuit.
It's like the difference between a game whose logic is 90% Lua or Python scripts, included plaintext in the game directory, vs. one that's 100% compiled C++. One is susceptible to modding by a 12 year old kid armed with a notepad, or a 22 year old kid trying to make a flashy visualization of finite state machines to get a good grade on CS labs for little work[0]. The other is... still mutable, if you get into reverse engineering, and probably pay for (or pirate) SoftICE[1]. More importantly, one lets you learn how to make similar things, through looking and experimenting; the other doesn't[2].
--
[0] - Well, that involved Processing to show an animated diagram of a simple FSM, and Colobot with a flying Moon robot programmed with that FSM for the flashy vis, plus some half-assed IPC using text files...
[1] - Ghirda wasn't a thing back then.
[2] - See also "Show source" in browsers - used to be a great on-ramp to webdev and programming, back when JS was just a toy language.
It's different in that analog circuits aren't just a physical representation of a wiring diagram - they are something I know I could manipulate if needed. Cut out a part there and replace it with a substitute, splice some wires in and add more parts to add extra functionality, etc. Digital electronics? The chips are magic black boxes. And when they communicate using digital protocols, like SPI, I2C, UART, etc., then the physical properties of connections between the chips become irrelevant too - the connection is either present or missing, you can't manipulate the behavior by messing with the wire from outside. No soldering in a cap to change delays, or a 555 to add timed behavior. Almost all of the circuit is therefore hidden in magic black boxes, and whatever remains either works or it doesn't.
Now sure, I know now I can do digital electronics with the right tools. But the tools I can use, that most hobbyists and professionals can access, allow only limited control. You can't just fab yourself an alternate chip to do something, you have to buy the very specific one that happens to do what you want, or a more general one that can be reprogrammed. And you rarely can reprogram existing chips on a board you modify, because the vendors don't want to see the magic.
Yes, I too, as a kid, was pissed off about digital electronics, and regularly remarked to a friend that soon we'll have light switches and fridges implemented using microcontrollers. If I only knew how true that prediction would turn out...
Now the irony is, at that same time I was making those complains, I was also studying X86 assembly and C++, trying to learn how to make video games. Turns out, the common thread that connects programming computers and analog electronics is accessibility. I could do both for almost free, while gradually learning through experimentation. I couldn't do that with digital electronics - the cost and educational barrier was just too great.
We have some sand to thank. Jokes aside, it is still mind boggling how most of the manufacturing is still dominated by very few companies, it's the perfect setup for WWIII
If the development of modern chiplet/stacked die is interesting, semiwiki also has interesting articles from some of the vendors including breakdowns of design and integration costs:
My take is that the rapid rise in heterogeneous solutions and complexity will provide some excess semi profitability, but at the cost of long run performance increases for "new nodes". Instead of one path forward there are now many.
Leading edge nodes are becoming harder and harder to develop on, particularly outside of memory and logic. This is because of limitations in terms of Ft, max voltage, layout, poor scaling of analog, etc.
Leading edge nodes nowadays generally release without a lot of key features and so can only be used for compute tiles, requiring other tiles for IO, analog circuits, etc.
As the nodes mature these features will generally get added on (that might not even be possible in the future if the current trend continues), at which point you can do a SoC with all these things integrated on one die (which was essentially the only option before we had the packaging technology for heterogeneous chiplet systems).
This is why heterogeneous chiplet designs are the future.
Source: my job is helping to develop and test analog and mixed signal circuits on leading edge nodes at Intel.
Totally agree, the complexity of 3D, heterogeneous nodes, and chiplet integration are necessary. Thank you for your efforts, and I wish you great success.
It's just that for over 50 years optimizing transistor pitch in 2D was sufficient to drive demand and investment. The clear winning single path forward provided the exponential growth we've come to expect. I suspect the complexity and uncertainty of this new heterogeneity does not. Underlying system and process simplicity and reliability support overlaying complexity in software and network.
I started in semi with 68020/30 to see the huge jump from 2um to 0.8um and stayed in on the trailing edge of analog down to 28nm. Maybe there is another 40 years of growth, but it feels like even though 3D stacking increases the power law of dimensionality, it may reduce iteration rate more. We've got 8 high HBM, but will we have 100 in even 8 years or 1000 in another 8 more?
> We've got 8 high HBM, but will we have 100 in even 8 years or 1000 in another 8 more?
No, probably not, but that's not the point.
The point is that now memory and logic (in case of HBM) can be further decoupled, allowing us to use processes optimized for each, instead of having to compromise one for the other.
This will allow for cheaper and faster iteration.
You're correct that 3D scaling will likely not go very far, but that's not the key part; the ability to tightly couple heterogeneous dies, allowing for specialization and decoupling of processes, is the key to further scaling.
Don't forget the cavernous structures which people are forbidden to enter, buzzing with the activity of objects controlled by those enchanted crystals, like the magic brooms in Fantasia, performing myriad arcane rituals with rare and exotic materials to make more enchanted crystals.
They perform elaborate cleansing rituals on any item brought into the crystal-writing room. When they enter, they wear special clothing, lest the impurities of the outside world taint the crystals.
Those magical tools being self-referential incantations written in the weird and decidedly non-human language of the crystals themselves, containing deep magic [0] derived from the secrets of life itself [1] exploited to solve problems of intractable complexity in creating the next set of incantations.
Disturbingly, using the magical tool twice on the same incantation never produces exactly the same result.
Those EUV lasers are themselves insanely crazy, and unlike the semiconductors they are part of producing, the EUV gear has lots of expensive short-lived consumables — like mirrors!
Yep. One thought experiment I like is how well I’d be able to carry forward human technological progress to some primitive group, given any descriptions or samples of said tech that I’d want.
Microprocessors are always the choke point, where I’d be hard pressed to reproduce one, and they form the basis for so much else.
The bootstrapping required to produce an iron tool from scratch (really, a small village with a surplus of food sufficient for one full time adult male + accessible ores) is doable in less than a lifetime with some recorded knowledge on how to do it. Even if they’re literally at the mud hut stage.
Producing even the simplest IC? Definitely not. And that is ignoring the need for electricity and everything else required to actually use it.
Coal is a recent "invention". The Brits were pretty damn lucky they found lots of it under the last tree they chopped down. It was all wood before that.
Microprocessors with tiny features are a chokepoint, but I suspect you could make a primitive one if you really worked at it, i.e. go back to 1960's technology, such as the AL1 or 4004.
Even going to 1925 and teaching how to make a MOSFET would help.
Batteries from wooden crates with newspapers, metal rods and is enough to do telegraph. That it is slightly harder ensures serious use.
If the drinking water is far away at the top of a mountain it isn't so convenient to throw your garbage in it, take a dump in it or float the dead in it.
nandgame and nand2tetris are important works of teaching for this reason. On the off chance I get thrown back in time to the exact right time for it to be useful, I'll be prepared!
Because if you need to explain GPS at the level of the impact of general relativity, my understanding is that by itself is already a topic normally introduced in the final year of a physics degree.
If you're OK with simplifying to "time passes at a different rate for the satellites, here's the equation, I will not explain why it works just roll with it", why insist on ceasing to use it if the foundation isn't understandable?
What you've suggested doesn't work at high resolution at sea or in the wilderness due to lack of radio towers at suitable frequencies, atmospheric variations and ducting influencing the signal of the existing systems.
Doesn't work in valleys or cities due to multipath reflections.
Only works in low rise areas or while flying at high altitude.
That's why the satellites got built in the first place, and why China and the EU added their own satellites to the US and Russian global navigation satellite system.
It is more funny than I initially thought: We keep making things more complicated and secret in order to make something else less complicated. Even the luddites accomplished only to make things more complicated.
I found a similar treasure some years ago: How we take convenience for self-evident. We love to pretend everything should be as easy as possible. You should really never have to think about anything, never have to learn anything, never do anything, nothing involving muscles, endurance or fine motor skills. Life is best without developing and without accomplishing.
Any human technology that the general population can't understand is an egregious failure in education that will be indistinguishable from large-scale cognitive impairment. -- h-d'a
There has never been a time when most technology in use was understood by the general population.
"The Last Man Who Knew Everything"[0] was 1773–1829 and the trend towards compulsory education was only beginning, 2 countries at the start of his life and 7 at the end[1].
It's like I keep saying: We etch runes into stones, and make the stones come alive and do our bidding by channelling lightning through the runes. And you say computers aren't magic?!
We are practically machine elves.. we enchant, transmute and bind mystical inscriptions onto crystals that are charged through an invisible and intangible energy to perform actions we could never do with our biological bodies. All of this is produced succinctly through an empirically pragmatic yet slightly esoteric process of a form of gnostic meditation we call the scientific method...
What is mystical to follow a procedure? And even breaking a big nut with a stone is an "action that we could never do with our biological bodies".
Currently it is more esoteric on how a baby (human but not only) is formed than how we build a microprocessor. So if anything I would say we are statistical acrobats, existing despite the numerous approximations in biology. Compared to us humans, microprocessors are predictable and boring.
The mysticism is an emergent property of sufficiently complex and "obfuscated" procedures... No intelligent entity can lay claim to an omnipotent and perfected understanding of all known procedures... The fragments between silos of rational derivations of existing predicated truths we have discovered about the natural order of the universe is where it feels more like an enchantment than a discovery.
I strongly disagree around how forming a baby is amore esoteric.. cell mitosis is a pretty well understood science at this point and eventually we will reach a hard limit of covering all the surface area of that domain of knowledge. However technological discoveries unfold more like a fractal.. it isn't really a bounded domain as far as we know.
You think so? For me personally, there's few things that are more inspiring than understanding something and being able to reproduce it. This for me is the wonder of science.
I've always thought of chips and the unrelated peter principle ("people rise to the level of their own incompetence").
Except with chips it is shrink instead of rise. Like if a chip works well and is reliable, time to shrink it (or run it faster, or run it hotter, etc) :)
From Monster Manual by Gary Gygaxale and David Ångströmson (TSR Inc., 1978), p. 61, in the section on etches:
"An etch exists because of its own desires and the use of powerful and arcane magic. The etch passes from a state of humanity to a non-human, non-living, and non-conductive existence through force of light. It retains this status by certain conjurations, enchantments, and a reticlefactory. An etch is most often encountered within its hidden chambers, this lair typically being in some dry, deserted area or vast underground lab, and in any case both solidly constructed of stone and very sterile.
Through the power which changes this creature from human to etch, the armor class becomes the equivalent of +1 plate armor and +1 shield {armor class 0). Similarly, cast dice are 8-sided, and the etch can be affecied only by magical attack forms or by monsters with magical properties or 6 or more hit dice.
Etches were formerly ultra powerful magic-users of magie-user/clerics of not less than 18th level of magic-use. Their touch is so cold as to cause 1-10 points of damage and paralyze opponents who fail to make their saving throw. The mere sight of an etch will cause any creature below 5 nm {or 5 hit dice) to flee in fear of overexposure. All etches are able to use magic appropriately at the level they had attained prior to becoming non-human.
An etch can only be permanently destroyed when their reticlefactory is destroyed. Unless the etch's reticlefactory is located and destroyed, the dice will be cast, and the etch will re-adhere in 1d10 days after their apparent defect.
The fallowing spells or attack farms have no effect an etches: charm, sleep, enfeeblement, annealing, insanity or death spells/symbols.
Description: An etch appears very much as does a wight or mummy, being of skeletal form, eyesockets mere black holes with radiating points of amplificated light, and garments most often rotting (but most rich)."
Packaging has become so advanced that they now need similar accuracy and clean rooms as the chipmaking itself. When they add microfluidics it becomes even more difficult.
So anyone want to hazard to guess if the node shrinks are going to go more than seven ignoring the half steps?
If the article is correct that the major semis are going to kind of forge their own paths, in my opinion, that means marketing lots of even 1/3 or quarter steps.
Intel is currently limping along on government handouts, hoping people don't notice their latest generations burn out if you leave them turned on. Taking their primary competition away would push up profit margins and push down the need to fix the engineering. If they aren't dead already, TSMC disappearing would do it.
I wonder how global foundaries is doing these days.
A specific fab had a process issue, that appears to have been rectified in April 2024. Intel 4 is not affected. Other fabs making Intel 7 Ultra is not affected. The issue only noticeably shows up in the highest-end processors, and dropping max clock multiplier by ~3% seems to fix things.
It's not a good thing, but it's not doomsday. The earliest Ryzens 1000s had issues with making incorrect calculations in certain circumstances, and look where they are now. More recently, just a year ago, certain Ryzen mobos literally fried their chips to ~1000 degrees (literally 1000 celsius).
Yes, there are reports of ~50% failure rates, etc... but if you get a shipment of a contaminated batch, it's probably gonna fail, and a single source shouldn't be used to generalise.
To be fair their "last" generation is already 2.5 or so generations old. Not necessarily a very accurate reflection of their current state, their next released will probably make or brake them.
> I wonder how global foundaries is doing these days.
I guess they are in an entirely different market than high-end Intel/TSMC/Samsung fabs?
The following in the article is pretty spot on ”Numerous industry sources say TSMC’s real strength is the ability to deliver process development kits for just about any process or package." There are countless flows and tools that need to get enabled, tested, etc. on the way to make a chip. TSMC has established a fairly good reputations delivering the required collateral. Just having a great process without proper PDKs will not get you anywhere fast
