Hacker News Reader

Comment by throwaway_pdp09 | Hacker News Reader

Why did they fail?

They are extremely secretive:

One of the funniest stories I heard was about the first copy of the Intel 14nm design rules Altera got from Intel. They were heavily redacted, which is something I had never seen in the foundry business.

https://semiwiki.com/semiconductor-manufacturers/intel/7912-...

Additionally, a contract fab tries to be flexible, easier to design for. They make their libraries & design rules work well with commercial design software. A private fab on the other hand can boost yield & performance by pushing increasingly complex design rules on the design team, and the design team can use customized or internally developed software to make it all work.

So in short, they had the best fab tech in the business but they weren't used to servicing anyone other than themselves.

Surely 'themselves' is a huge enough market, and 'boost[ing] yield & performance' is exactly what they want to do, as the world's largest CPU maker? It sounds perfect. I guess maybe financially the extra perf etc. just wasn't worth it.

(thanks for answer though)

There is no reason why you cant optimise for both. To simply put they wasn't ready to be a contract manufacture, and TSMC set the standard of what is expected to be a contract manufacture in the Fabs industry and Intel failed miserably to meet even the pass mark.

Yes, it make perfect sense to optimise for you own yield and performance until TSMC overtake you and you no longer has the volume advantage. In manufacturing volume is everything.

It is also worth pointing out Altera is actually a Subsidiary of Intel. If that is what they were going through you can imagine about the others.

Maybe they do know how to optimize, but not exactly why anymore?

I've seen a tv-documentary about 15 years ago about their fabs. Therein some reporter asked why there was a bend in some bundle of pipes for no apparent reason. They answered that was there because in another factory there was a column there which they had to route around, which didn't exist at this factory. They explained further to not wanting to risk process variation, by building that bundle of pipes straight again. Since then i have been unimpressed by all that high-tech, being so complex, having to cargo-cult around, instead of being able to understand it from first principles and build according to the situation. Not blindly copying something which worked at another place but doesn't apply anymore, because the place and its conditions changed.

I don't think it was cargo-culting in the case of that fab. A fab isn't like your mother's ham, which she cut in half because her mother cut it in half, only to find out that her mother had a small oven. A fab is like that tricky piece of code that has a comment on every line, but instead of one function it's 20,000 lines of that. On top of that, it takes several weeks before you know if your wafer came out okay. So when you duplicate your fab, the last thing you want to do is add more potential problems. Think of it like using a software library; you don't go rebuild the library each time you use it, you just copy it into your project. Sure, there's unnecessary piping, but you don't care. If you cared you'd write your own library in assembly language and it would be a shining example of efficiency and understanding. But maybe you just want to solve your problem and ship.

Yes, that way we got fragile bloat, eliminating all benefits of every new generation of hardware. Since the 70ies. How long will this model of 'value-added' innovation be sustainable, i wonder?

I used to work in R&D for developing the machines used in the fabs. Sometimes the trivial details make a difference in the results and you don't always have to time to isolate and analyze each little thing. If you have a proven recipe, don't change it unless you have a reason.

I understand what you mean, time to market, pressure, deadline, and so on...

But see where it got them?

I think your criticism is cheap and unwarranted. Where you see stupidity I see wisdom, though we can both agree it's ugly, but let's ask a straight question.

If it were your $$$billions being spent on a fab, what would you do?

Probably going broke because of unwarranted perfectionism.

/me lifts hat and bows.

Consider the compliment well and truly returned for your honesty - thanks!

While sleeping and cuddling with my cats, they whispered into my dreams: Stoopid hooman, alwayz so slooow! If yooo'd got Billionz to burn, yoo'd push forward minimal.fab of Yokogwawa and BiZEN from Searchforthenext/Wafertrain! Rrrr Rrrr Rrrr!

Courtesy of Lilly, Mieze, and Mimi.

When I was in school for Microelectronics, we were told very early on that experiments needed to be designed for the least number of runs, because of the sheer cost of taking a production fab down for a short period. That and the design better be flawless or your manager would never allow it

To a first-order approximation, engineering is about minimizing costs, and a large fraction of the cost of creating and running a fab is the cost of the time of the employees able to understand things from first principles. So, just because they decided to avoid that cost doesn't mean they are unable to redesign the pipes from first principles.

The point is that they probably have no idea what else they would have to change/retune in that process to account for the pipes being straight. It makes sense that they would seek to avoid changes to the greatest possible extent, given the costs involved.

How many R&D engineering hours do you think the extra 1m of piping equals in cost?

I'm actually surprised they didn't go and put a column there, too, just in case.

You're describing Copy Exactly (CE!):

https://www.extremetech.com/computing/127987-deliberate-exce...

> Intel began working on CE! after it had trouble at the 0.5um (500nm) node and refined the process through each successive generation. The green line shows the initial product ramp at its first fab — after an early spike, yields cratered and only recovered over a period of months. Once Intel had Fab 1 working well on 0.5um, it started ramping Fab 2 only to run into new problems. As Copy Exactly! was developed and deployed, the company’s yields synchronized across the various fabs.

Because they don't know enough to predict the process effects of variations in humidity, barometric pressure — or pipe length! — they eliminate those variations instead. Presumably they would learn more if they didn't do that, but they would also take longer to ramp up production. Since they started doing CE! in the 1980s, it's probably not a primary cause of their late-2010s decline.

Altera was already working with Intel before they were acquired.

Well, exactly, they are their own best customer. Contract fab was just a way to fill capacity in a lull period. So there was never a strong motivator to make sacrifices to build a great contract fab program.

I guess the software analog of this is that the quality of code usually goes up when it’s open-sourced. Not just because of the contributions from outsiders, but because the original developers feel apprehensive about putting out low quality code.

For a contract to be mutually profitable it has to somehow be better than off-the-shelf. I got to imagine coordinating custom chip development is expensive for both parties.

Reply on news.ycombinator.com