Feels like we are quickly centralizing consumer chip fabrication into a single company. I guess the barrier for entry is so high and TSMC is just so far ahead.
TSMC only just pulled out in front. Rewind the clock a mere 5 years and Intel was in front, with Samsung and GlobalFoundries basically tied for 2nd, and TSMC in dead last (they had the weakest 16nm/12nm of that generation - the only one who couldn't hit 30MTr/mm2 of the bunch)
GloFlo then backed out entirely of the race and Intel slammed into a wall.
Since then Samsung and TSMC were on "equal" ground at "10nm" (both ~52MTr/mm2, both released 2017) and again at "7nm" (both ~96MTr/mm2). It's not until 5nm that TSMC was actually clearly in-front of everyone else, with their 5nm being 173MTr/mm2 while Samsung's is only 127MTr/mm2.
In terms of "TSMC is just so far ahead." Samsung's 3nm is supposed to use GAAFET while TSMC's 3nm will still be FinFET. So.. potentially Samsung re-claims the "crown" so to speak at 3nm. And Samsung does contract out their fabs - see Nvidia's RTX 3000 series. There's also no particular reason to believe that Intel is down for the count for good. They are a huge company with a huge amount of capital, they can fund a rough generation or two.
I remember the huge controversy around the iPhone 6S which had a processor that was either a 14nm Samsung or 16nm TSMC (https://en.wikipedia.org/wiki/Apple_A9). Apple had used Samsung for the iPhone 4 through iPhone 5S. The iPhone 6 was TSMC and the 6S was dual-sourced (20015/2016, 5 years ago).
I'd guess that some of TSMC's rise to prominence was partly driven by Apple not wanting to help Samsung. Apple was really pissed about Samsung copying the iPhone - not just by shipping an Android phone, but by copying icons to make their phones seem as similar as possible.
You can definitely look back at MacRumors articles from 2014 and see that there's a bunch of Samsung/GlobalFoundries/TSMC "who will be able to make it happen" talk.
In fact, re-reading through these articles, it seems that people thought that Samsung and its alliance with GlobalFoundries would be the winner of the iPhone 6S generation, but it's possible that Apple saw better yields from TSMC and saw the potential there. When you're as big as Apple, you're going to be really deep with your suppliers and you're going to have a lot of expertise to judge suppliers and their future potential. Maybe it was a combination of seeing TSMC over the iPhone 6/6S generation that gave Apple the confidence to move away from Samsung. Back in 2015, analysts were still expecting Samsung to be getting future business from Apple like the A10X processor.
Given that Apple is a buyer that can move mountains, how much of TSMC's ascendency is potentially Apple committing to a lucrative multi-year deal allowing TSMC to invest a lot of money knowing they had guaranteed orders? One of the hard things in business is knowing what to spend your time on - what do customers really want. Google, for example, has spent plenty of time on things that weren't good investments whether that's Wave or AppEngine or Google+. If you know "doing X will definitely make me a lot of money" it makes it easier to invest heavily in an area - basically, you kinda get the benefit of hindsight ahead of time with a long-term deal.
I hope Intel and Samsung continue to do well (or get back into the race as Intel's position might be) since more competition means lower-cost processors over the long-run. But I think it's definitely important that you point out that only a few years ago TSMC wasn't the powerhouse it is today. While I believe TSMC is going to continue to invest and improve, Samsung is producing Qualcomm's Snapdragon 888 on its 5nm process and if you're right about Samsung's 3nm process, that should provide a lot of orders there too - especially if Intel is willing to outsource manufacturing.
The job of a gate is to "hold open" / "pinch closed" a channel with an electric field. The closer to the gate, the better the hold/pinch. In ye olde days, you'd slap a gate on the top of a channel and call it a day. Every part of the channel was close enough to the gate get a good pinch.
Then everything shrunk and smaller channels wound up needing stronger pinches to completely shut them off. Instead of slapping a gate on top and calling it a day, they raised the channel into a fin and drizzled the gate over 3 sides so that it could pinch from the left and right, not just the top. Those are FinFETs.
The next step is to have the gate on the bottom, too, so that it can pinch from all four sides. The channel literally goes through the gate, which surrounds it on all sides. Those are Gate-All-Around FETs, or GAAFETs.
This is beautifully written; I stepped away from keeping up an in-depth understanding of silicon processing a bit before FinFETs took over and this feels like one of the best introductions to the basics of gate geometry of FETs.
I never quite got how electrostatic control was supposed to work with GAA, Fin and derivatives: without the bulk, how is your electric field supposed to work?
Unless the Gate itself is hollow, and the "bulk" is in the middle, like air would be in a hollow spaghetti? Or tha you use way higher voltages for the gate than for the drain/source (but then, how do you drive that?). I guess I just need to look it up, but most presentations gloss over that part.
In the 90s there were dozens of leading edge fabs. But while the barriers were high back then the capital investment needed to get into the next node has gone up exponentially, about 15% each node, since then doubling every 5 years. It took less than $1 billion to get in the game back then but over $20 billion now.
It's not clear whether this is cause or effect. The insatiable demand for silicon fabbing has arguably made a $20B plant today more economical than a $1B plant 25 years ago.
What’s tricky to understand is that once TSMC said, ‘we will build everything and design nothing’, everyone looked to them when there wasn’t another competitor willing to do the same.
The massive order volume they received (with low margin) let them experiment on process development 10x that of Intel.
Remember when Wall St. said America can outsource manufacturing because ‘other people’ aren’t smart enough to innovate? Looks like they were wrong :)
The American middle class is going to be further decimated over the next decade. Wups
On another side - has the software side done much to incentivize Intel to innovate over the interim? i.e. Windows has accrued a lot of cruft. It feels pretty outdated at times. It feels almost unprofessional at times coming from a unix-like OS user space and has since the 90s. But then unix-like OSes have been a thing this whole time.
Yeah, I think Intel is our latest, greatest example of the follies of not regulating markets properly. Properly regulated, Intel would have been smacked around or incentivized against dominating the chip market 'back when'. Given the factors required, this may have had to require state-led/funded chip research and fab production. i.e. Effectively making a government do what TSMC did years ago - and privatize the actual results of the efforts at key points/areas. Just enough public investment to push the market toward efficiency then back out. Proper regulation would have kept pressures in place to keep key technologies and manufacturing processes/abilities from entirely leaving the local market, too.
99% of the barrier is feature size. Producing the feature size starts with photolithography.
If you want to talk about centralizing concerns, look into the number of companies who can produce an EUV light source capable of supplying a photo tool with powerful & precise output 24/7/365.
The main problem is that there is no profit in chip manufacturing ( relative to software development). Apparently it takes 5-6 years for a node to make any profit.
> The main problem is that there is no profit in chip manufacturing
This isn't true. Check out the margins of TSMC, it tends to be 40% or so.
But to maintain this lead in the industry, they need to massively reinvest for the next smaller process. With that said, profits are great but they don't endure (software development is somewhat more 'sticky' especially since everyone is doing SaaS which provides more incentives for competition and many companies are growing even with covid19 changing the market landscape).
Maybe it is so with the monopolized ( or duopolized if we consider Samsung) foundry today. I visited Global Foundries just before they stopped investing in smaller nodes ( and Infineon which was nearby). Listening to the profits they made and the problems they had ( and my own experiences), I realised that the complete business flow for chip manufacturing is flawed somewhere. I did not have enough motivation to look into detail at that time.
Of course, most of the big foundries would still be open, if the profits were anywhere more than 10%.
I'd say it's weird because hardware like this is so much harder, but the benefit's also a lot more marginal. There was a time when you had to buy a new computer every 4 years or it would be cripplingly slow. These days, pretty good hardware that's 8 years old is good enough today if you have an SSD, 8GB of ram, and don't play AAA games.
There are a lot of other innovations possible, but the business flow for hardware manufacturing does not have the right motivations. I personally think that Apple is doing the right things on several levels.
While I understand the concept of "technology node" as a manufacturing process, where does the usage of the word "node" come from? From litography? Is it related to the "nodes" in the electric circuits?
> The technology node (also process node, process technology or simply node) refers to a specific semiconductor manufacturing process and its design rules. Different nodes often imply different circuit generations and architectures. Generally, the smaller the technology node means the smaller the feature size, producing smaller transistors which are both faster and more power-efficient. Historically, the process node name referred to a number of different features of a transistor including the gate length as well as M1 half-pitch. Most recently, due to various marketing and discrepancies among foundries, the number itself has lost the exact meaning it once held. Recent technology nodes such as 22 nm, 16 nm, 14 nm, and 10 nm refer purely to a specific generation of chips made in a particular technology. It does not correspond to any gate length or half pitch. Nevertheless, the name convention has stuck and it's what the leading foundries call their nodes.
