Hacker News Reader

> assets from IBM, which GloFo "acquired" in 2014 (for -1.5 billion at the time)

You used the term "acquired" along with a minus sign in front of 1.5 billion but just in case other people don't get the joke, IBM gave their fabs to Global Foundries and PAID them $1.5 billion to take them.

For what it's worth, IBM is suing to try to get its $1.5 billion back, claiming that GlobalFoundries didn't uphold terms of the deal: https://www.burlingtonfreepress.com/story/news/2021/06/11/ib...

Intel could fix that. Could be an added incentive for the deal.

That is incredible. Why did GloFo think they could just take the 1.5B and not deliver? I guess if they don't have to pay too much damages it's a huge free loan...

This is America and they’re a corporation. It’s the price of doing business.

Because example has shown time and again that you can take the $1500 million, spend $100 million on legal fees, settle for $400 million and keep a nice $billion

Their interpretation is probably that they, as promised, spent the 1.5B trying to deliver and it didn't work out, too bad.

Intel PowerPC. Who would have thought it would happen?

:-)

Yep. I'm glad others are aware of this. There were certain manufacturing processes that needed to be sustained for "national security" reasons. GF was an imperfect solution but it was better than the alternatives at the time.

Intel did offer foundry services, at least 3 years ago or so, albeit designing stuff for that node required putting everything in separated environment, and simulating in separate compute farms. Even TSMC which was and is far ahead does not demand this level of paranoia, only documentation is put in secure isolated environment, other stuff being managed in the more reasonable way of unix groups, which are, well, good enough.

Isn't the difference that Intel is actually protecting its chip IP, whereas TSMC just fabs other peoples' chips and isn't in the business themselves?

Just designing on their node as a foundry customer.

Sounds like they need some Tom Smykowski's ...

Anyway I wonder if GlobalFoundries will be it's own business unit or otherwise get enough independence to do things the right way / differently than in the past? I often find large companies who are 'not good with customers' aren't really great at acquiring competency at being good with customers if they just buy and assimilate it.

Heck I've been a part of companies who had a big customer focus and they acquired smaller companies who weren't good with customers, and somehow the smaller company infected the organization and suddenly they became bad with customers :(

LOTS of companies are either suing or thinking of suing GF for failure to uphold technology development/adoption agreements. IBM is one of those.

This Intel situation is certainly part of "keeping it in the USA" though it would have been far easier and cheaper to prevent outsourcing back in the mid-1990s when Silicon Valley's silicon ebbed out of the country. It was a night-and-day change and lots of knowledgeable people were scattered to the wind back then. Now much of that knowledge is having to be reinvented and rediscovered from scratch (because it was tacit knowledge that never could be recorded into books or other writing).

This probably is the best way for Intel to become a foundry. They tried 10 years ago and the entire process was 100% Epic Fail. Not surprising given the pain seen when AMD split off its fabs to form GF, when it abandoned being an IDM vertical fab. When you are an IDM, many process-design interfaces break or have to be formalized in ways that were never imagined or learned by the organization. That's how Intel failed its foundry attempt before so simply avoid Intel culture by buying an already working culture makes a lot of sense.

> (GloFo licensed their 14nm from Samsung as a reminder after their own 14 failed, and their 7nm failed too)

Can someone with insight explain how this can happen? GloFo is a company worth billions, how can they just fail at this while Samsung succeeds. What's so hard about this in particular? I don't think I recall companies at this scale just failing at something that is supposed to be their area of expertise.

I‘m looking forward to historians writing a big post mortem about this in a few decades. For an outsider it is really hard to get to the bottom of what makes a company succeed or fail in this space. It seems like „throwing money at it“ is reaching its limits in this case

Considering they do not have the compensation reputation that FAANG have, I would not conclude they are “throwing money at it”.

I think the charitable answer is that bleeding edge process development is getting very hard and many exited/stumbled. IBM failed at 14nm too, and Intel which famously called itself 2 years in front of the rest of the industry for decades has been stumbling hard since 2015 and still hasn't really replaced their 14nm on the most visible parts of their lineup.

As to what failing means, it can be many things, sometimes you just have way too many defects to have the process be commercially viable (if your cost is 10x your competition and you can't fix it, you don't have a process), some bad technical choices because there's been less and less cross industry cooperation (things like the ITRS that gave a rough roadmap hasn't been a thing for a while, last I checked), and introducing new technologies (EUV or new metals) may compound those. Sometimes chips in your new process are more dense (which is good), but perform poorly because of one of the previously mentioned stuff to the point it's hard selling new chips that perform worse than previous gen. In Intel's 10nm case there was a bit of everything and it will be very interesting to know exactly what factored in more (Cobalt is often rumoured, a very overconfident choice for MPP was my initial understanding, I haven't really followed on this).

In the case of GloFo, it's perhaps simpler than that : GloFo was part since 2007 of an alliance between them (initially as AMD), IBM and Samsung called the Common Platform [1] that had more or less standardised a process, with a lot of the heavy research done by IBM (I would say, out of spite for Intel) and then used by GloFo and Samsung.

14nm was the first that everyone was on their own, because IBM failed their research. In the end, Samsung managed to get it done, and GloFo either failed or... didn't try really hard ? This is where I'm gonna speculate a tiny bit and be less charitable, I don't think that GloFo was sufficiently funded in terms of process development, despite their parent owner being able to.

And it definitely looked like from the outside, at least from my point of view, they were always in dire need of money except for acquisitions, and late with everything (including the process inherited from the alliance). Most of the business decisions done (including buying Chartered and reverse buying IBM's 22nm fabs) seemed to be about pure business and missed the technical understanding of the business and what it required to keep it going.

At the end of the day, for them, licensing 14 from Samsung was a great out of that mess. But that didn't happen for 10, and we (at least I) don't really know why. My guess is that GloFo would probably have been happy to keep going on with that arrangement, but I may be completely wrong on that one.

At the time I seem to recall that both had some issues filling up their order book (which seems a bit insane today, considering the current shortage) and that may have played into it from Samsung's point of view.

[1](https://www.chiphistory.org/546-ibm-microelectronics-common-...)

> 14nm was the first that everyone was on their own, because IBM failed their research. In the end, Samsung managed to get it done, and GloFo either failed or... didn't try really hard ? This is where I'm gonna speculate a tiny bit and be less charitable, I don't think that GloFo was sufficiently funded in terms of process development, despite their parent owner being able to.

For whatever it's worth, it feels like GloFo has had various process issues for a -long- time. While there were issues with the Bulldozer design itself from a deep pipeline perspective etc, the other factor in it's lukewarm lifetime (especially the first couple gens) was issues with GloFo's processes even back then. I know some of that was that they were trying to gear more towards bulk silicon vs CPUs but I think there were other issues too.

> While there were issues with the Bulldozer design itself from a deep pipeline perspective etc, the other factor in it's lukewarm lifetime (especially the first couple gens) was issues with GloFo's processes even back then.

I think no process could have saved Bulldozer for AMD, it truly was a deeply flawed design IMO, but yep GloFo certainly didn't help. I think my "uncharitable" take was a bit too kind considering.

Specifically in the case of AMD, GloFo relied early on on guaranteed income from AMD through the Wafer Supply Agreement which was stupidly tight (my understanding is that it was part of the deal of selling to Mudabala, though exact details are not public), included a massive volume of must buy from AMD, and didn't include any penalties for GloFo missing deadlines or having their processes underperforming. They definitely coasted on this and I should probably have mentioned this above, since it led to many of their issues going forward till 14nm.

It took them years to manage to loosen the WSA up, the last step was (if I recall correctly ?) loosening the CPU core exclusive that I believe was part of it (they had a carve out for GPUs which were built at TSMC at the time of the sale, though that was supposed to be phased out if my memory is correct, before it was renegociated and AMD moved the low end GPUs there instead, except maybe one gen?).

To their credit, AMD pretty much never bad mouthed GloFo in public. Managing to do the original Ryzen at GloFo and getting as much out of the 14 process that they did is, I think, one of the most underrated engineering tour de force from those who worked on that. While supposedly "copied exact" from Samsung, it wasn't exactly fully there at the time.

I did read that the last renegotiated WSA between AMD and GloFo is supposed to be going on till 2024 (?) currently, so if this goes through, I hope there's a "in case of sale" carve out, if just so anyone can save face a bit ;)

(and that would be a pretty large unburden from AMD in any case, which makes the whole thing even more ironic)

Yes, I was surprised as well. That Global Foundries doesn't seem to know about this suggests to me some shenanigans on the street.

However, if they did buy GloFo it would give them a straight forward path to having a "Fab for hire" service ala TSMC and their own fabs as well. Kind of a have your cake and eat it too thing if you can bring GloFo's ability to bake chips up to the requisite standard.

No, Global foundries doesnt have a viable 5nm or 3nm node yet, yeah they have suspended all 7nm operations and researching on creating the 5nm node, but it will take time. It makes more sense to buy TSMC than GloFo.

Not all chip foundry customers want or need the latest greatest nodes. We can see this in the current chip shortages, all those dozens of embedded CPUs needed in cars, and in other consumer devices aren't sub-10nm designs. The vast majority of chip manufacturing is actually at fairly pedestrian nodes for cost reasons.

Maybe TSMC isn't for sale.

TSMC's market cap is more than 2x Intel's

So then have tsmc buy intel???

The US would block that acquisition. No way they'd let Intel go to a foreign country.

I sincerely doubt the US Government would allow such an acquisition. The only way I could see it happening would be a monumental failure of communication between DoD and the FTC. DoD absolutely requires a source of chips designed and fabbed on US soil. As long as they are aware of anything going on, they'll fight tooth-and-nail to keep Intel independent.

This is such a big deal that it presents problems for vendors working on even completely open, non-sensitive systems. Like my company. We are a foreign language instruction company that serves primarily DoD. Some of our customers are being super reluctant to approve a specific project of ours because it is nearly impossible to buy appropriate hardware that is not manufactured in China. We have to settle for a hardware vendor that is not our favorite choice because at least the design company is US owned.

It's not technically correct. It should be fine to use these systems for non-sensitive purposes, but you try negotiating a contract with a CO who won't explain themselves/refuses to clarify with his own command that this is OK.

The DoD probably wouldn't even have to get the FTC involved (directly), they have enough leverage in CFIUS to stop it there due to fabs being Critical Technology.

Correct. CFIUS swings long and low.

I am a bit interested in the relationship with Abu Dhabi.

You hack from the weakest point but unimportant part. You start from area not directly related …

Well yeah, that's what we're doing. We have enough other business to build the product on our own. But it's slower. Access to capital and all.

Not just buying their old fabs, I believe GlobalFoundries is still AMD's main foundry, so most of AMD's products would be built by Intel. The 'Intel Inside' stickers are going to get very confusing!

> I believe GlobalFoundries is still AMD's main foundry,

Only the IOD are still on GF 14 / 12 nm, along with some APU. Everything else are TSMC.

I do wonder if this wouldn't be an antitrust situation.

Why would intel need patents??

Theoretically, no two patents overlap, otherwise they'd infringe on each other. So any newly acquired patent covers something every existing patent does not, including your own portfolio's coverage.

Well, actually it's perfectly possible for an older patent to be broader than a newer and necessarily narrower patent. That is in fact the goal of writing a valuable patent. Even if only a narrow thing has already been claimed, but the spec enables something broader then only something narrower still can be patented after that publishes... otherwise it is not novel.

Of course not all prior art is in patents so any journal publication could also force narrowing of claims. Older patents are almost always broader... if they cover the thing you want at all.

AFAIK overlapping patents are fine and common. It just means that when implementing the technology you need to license/own both patents.

Isn't that only the case when the patents cover two different parts of your own implementation? How could someone patent something that has already been patented?

Team A creates a patent covering X. Team B licenses Team As patent and creates a patent covering X+A.

Turns out everyone who does X also needs to do A, to the extent X is X+A, so everyone needs to license both patents.

Makes sense but Team B's patent would be for their work, not for Team A's work, right? I get that you'd need to license both in this scenario, but the actual tech claimed by the patents wouldn't overlap.

Let’s say you invent the cart, and in the patent, you describe a platform with two wheels and two handles to push it.

Then, I can patent a cart with “a number of wheels” and “at least one handle”.

That’s why patents tend to be so vague.

You don’t want to keep the tiniest door open for such cases, so you don’t use a mirror, you use a light-reflecting device. If you’re smart, you don’t even use the term ‘light’. Your invention likely works with radar and röntgen, too. Maybe avoid electromagnetic fields, too. Who knows whether something similar, in the abstract, is possible with sound?

You also don’t say a device needs to do A, then B if there’s a tiny chance it might be possible to accomplish the same doing B first, then A.

This sounds horrible. Isn't the patent is meant to be practically usefull by a proffesional to build the device?

From my experience reading US patents/applications, they aren't just vague like that. They start out specific and then add additional claims which expand it. So they will have some claim about a cart with a specific number of wheels, but in later claims they'll add language to generalize so any number of wheels are covered. There are also figures and detailed descriptions which are separate from the claims which aim to make clearer. Pretty much any one I have read seems to make it clear enough create the device.

That's not really a good example since a cart with two wheels and a two handles anticipates under 35 U.S.C Sec. 102 a cart with a number of wheels and at least one handle.

But yea generally you want to describe things with a billion examples, but no statements limiting the scope of your invention.

The most devious strategy is that you disclose vague and probably impossible embodiment. Wait 10 years and then write claims that that narrowly target big products that you really couldn't have invented ten years. But the burden is more or less on the accused infringer to provide you didn't describe the new invention in the old patent.

> But perhaps more importantly, I think that what Intel is buying is what was sorely lacking them : people and a methodology of working with external customers. From all I've heard over the years, Intel's process and design teams are so insular and out of touch with the rest of the industry's practices that they had a terrible time just working with customers.

This is not a problem at all. If they wanted to go this direction it's as simple as dumbing down, cheapening their process a bit, and giving free reign to make decisions on what they will be fabbing to technically competent salespeople.

With Intel's resources they could've even developed such commodity process separately to their own process, and ran them in parallel.

They had almost a whole decade to do that. It's just they simply didn't.

Simple? When you have culture issues as Intel does, there's nothing simple about any of this.

> It's just they simply didn't.

They've wanted to go in this direction for many years and so far failed. We've had meetings with them over the years and both times their offering was uncompetitive. I'm not sure what the sales people are supposed to do if the core technologies fabless customers want are not available.

I lot of people don't realise Intel are already in the foundary business. I'm not surprised their customers are not shouting about it - why would you want to tell everyone you are using an arguably failed process.