IBM Versus GlobalFoundries: A Lawsuit Instead Of The Power Chips Planned - IT Jungle

kuaciasing.blogspot.com

June 14, 2021 Timothy Prickett Morgan

Last week, IBM took its former foundry partner, GlobalFoundries, to court in a lawsuit that alleges, in essence, that the company promised to deliver Power9 processors based on 14 nanometer technologies and Power10 processors based on 10 nanometer technologies and had some issues with the former and never delivered on the latter.

Not only that, IBM’s lawsuit says that it never released GlobalFoundries from its from its promise to deliver a 10 nanometer chip, and when the company promised to shift to 7 nanometer technologies, IBM in good faith worked on developing Power10 chips for these processes and spent at least $188 million. And further, in the summer of 2018 the foundry – which was at the time an amalgamation of the old foundry operations of chip maker AMD, the IBM Microelectronics division it acquired, and the chip making operations of Chartered Semiconductor – decided not to go ahead with 7 nanometer process development at its Fab 8 plant in Malta, New York, where the Power9 chip and the System z14 processors based on 14 nanometer technologies are made.

GlobalFoundries stopped development on both standard water immersion lithography and more advanced (and difficult) extreme ultraviolet (EUV) processes, and that left IBM in the lurch for Power10 and presumably for Power11 and Power12, which would have been delivered in the 10-year period that the two decided to work together over when they inked their partnership in July 2015.

The lawsuit comes at a peculiar time, with Big Blue getting ready to ship Power10 chips in high-end NUMA machines starting in the third quarter of this year and GlobalFoundries apparently getting ready to do an initial public offering in a world that is hungry for semiconductors.

These Power10 chips, as you know, will be etched by Samsung Electronics, which wants to expand its fabs beyond making CPUs for client and mobile devices as well as main memory and flash to making high-end server parts – perhaps some day its own Arm-based server processors. The word we hear is that IBM waited to sue GlobalFoundries because it did not want to jeopardize its supplies of Power9 and z14 chips and therefore sales and profits at its Systems group, which has been utterly dependent on the CPUs coming out of the Fab 8 facility of GlobalFoundries.

I got my hands on the documents and wrote up a little history of the relationship between IBM and GlobalFoundries and my analysis of the IBM suit and the initial GlobalFoundries response, which you can read over at The Next Platform at this link. I feel no need to repeat everything here, but I do want to focus in on the effect that the difficulties GlobalFoundries had in making Power9 and Power10 chips has had on IBM’s Power Systems business and for IBM i shops in particular. I will assume you have read this analysis as background so we can talk about these effects. (Not to give you a homework assignment offsite.)

The simple fact is that IBM reached a point where its own scale at its East Fishkill foundry had hit its economic limits. Running a modern foundry is a very expensive endeavor – it costs around $10 billion to make a foundry of reasonable capacity that can etch 7 nanometer chips – and IBM simply did not have enough volumes of its own and enough merchant deals for its foundry to stay in the game. Particularly after losing all of the game console processor deals with Microsoft, Sony, and Nintendo. GlobalFoundries absolutely had aspirations to compete against Taiwan Semiconductor Manufacturing Corp, which makes all of the most advanced chips in the world – these days etched using 7 nanometer and 5 nanometer processes – at this time from its foundries in South Korea and China. I know because I visited the Malta Fab 8 foundry in early 2018, which was apparently when things were, behind the scenes, going badly between IBM and GlobalFoundries.

What seems clear in hindsight and from changes in the Power8, Power9, and Power10 roadmap is that this partnership had trouble from the beginning and IBM’s own foundry appeared to be having trouble, too. I don’t think that IBM ever really wanted a three-year or four-year cadence between Power processor updates – that is simply too long to remain competitive, even if IBM has been able to make very large leaps in performance over those gaps that make it look like it is still on the Moore’s Law track. Customers expect smoother curves and regular price/performance enhancements, which is why we see AMD with Epyc X86 processors and Ampere Computing with Altra Arm processors making promises for annual updates to their server CPU lines. That may be too much for the enterprise Power Systems and System z customer base, but I think a two-year cadence is reasonable. A processor launch – Power4, Power5, Power6, and Power7 – was the first iteration, followed by a refinement in chip manufacturing process or a whole new process, when possible, with an updated generation – Power4+, Power5+, Power6+, Power7+, and so on – as the second iteration. This didn’t always work out, of course. Power4+ was not much, and IBM didn’t even call Power6+ by that name until after the fact after I caught them rolling out new machines with the chip and not telling people that.

Now, IBM is in a much more risky position where it is changing chip architectures and manufacturing processes at the same time – ticking and tocking at the same time, in the Intel parlance – and that is just cuckoo. (To stick with the clock metaphor.) IBM did not do a full-on Power8+ chip, which we think it should have done for the HPC and AI community, and that might have had something to do with yields on its 22 nanometer process in its East Fishkill fab as well as the relatively limited number of potential customers that might want an enhanced Power8+ chip. So it did some half-assed Power8’ chip in 22 nanometers as a prototype for certain HPC customers who were waiting for Power9. I would be willing to bet that the plan was to get 14 nanometer for this Power8+ and actually try to make some money on it. Intel had 14 nanometer chips in the field in the fourth quarter of 2014. I see no reason why IBM could not have had 14 nanometer chips in the field in 2015 or 2016.

When Power9 did finally get announced in December 2017 using the GlobalFoundries 14 nanometer process in the Power AC922 systems used in the “Summit” supercomputer at Oak Ridge National Laboratory and the “Sierra” supercomputer at Lawrence Livermore National Laboratory, this was later than expected and had its own issues, according to IBM, which helped out because Big Blue was working on its own 14 nanometer technologies before it inked the deal with GlobalFoundries to offload its Microelectronics division. By the time the Power9 chips were rolled out across the Power Systems line throughout 2018, it had been four years since IBM put a new processor into enterprise-grade Power Systems. And I strongly suspect that the plan was for 14 nanometer processes to be refined and yielding highly so that Power9 would be “cheap as chips,” as they say in England when they refer to what we call French fries.

All foundries have issues and all chip designs have flaws, so it would be very difficult to argue that if mistakes are made, it is a criminal offense. The whole chip business is the triumph of optimism over the limits of chemistry and physics. It is truly magical that any of this works at all, to be honest, and it is a testament to the incremental and massively parallel steps that human beings across time and space can make to create true leaps in technology. If mistakes were a criminal offense, chip divisions of companies would be suing fab divisions – when they had them. This is difficult stuff. That said, delays in manufacturing processes and flaws in chips do cause economic damage, and IBM will probably be able to argue that it was damaged by the issues it had with GlobalFoundries. And GlobalFoundries will be able to argue that this is normal. And a jury that is almost certainly unacquainted with the eccentricities of the chip business will decide, and that is never a good place to be.

As far as I can tell, Power6+ was supposed to be a shrink to 45 nanometers, but it was a tweak to the microarchitecture and the doubling up of pairs of chiplets into a Power6 socket, just like the Power5+ shrink to 90 nanometers in the chart above allowed for doubled up chips in a single socket and Power7+ also used pairs of four-core chiplets to create much cheaper eight-core chips than the original monolithic eight-core Power7. IBM has been at this chiplet game that all the cool kids are now talking about for a long, long time.

That’s why I think the original plan – and the one that got screwed up by GlobalFoundries pulling the plug on its 10 nanometer process – was for a Power9+ to be etched using 10 nanometer processes and for a pair of twelve-core chiplets to be manufactured using that process. And then Power10 could be a pair of 24-core chiplets at SMT4 and a pair of 12-core chiplets at SMT8 etched in a refined 10 nanometer process, and for that to come out in 2020 – and I mean early 2020 for enterprise systems and late 2019 for exascale-class supercomputers funded by the US Department of Energy. That obviously did not happen, and for a lot of reasons and not all of them due to GlobalFoundries changing its process roadmap and switching it up to 7 nanometers. Somewhere along the way, when all this was going down – or rather, not going down – IBM just threw its arms up in the air and threw the “plus” generations of Power chips under the bus.

That was the wrong move from a market and competitive standpoint, but it probably was one IBM had to make given the constraints of its foundry partner. And those same constraints would be have been worse if IBM Microelectronics – which was hemorrhaging money in 2013 and 2014 before GlobalFoundries took it over and received a promise of $1.5 billion from IBM if it invested in 14 nanometer and 10 nanometer technologies over the decade between 2015 and 2025. It is important to note that 7 nanometer technologies were never mentioned publicly.

In any event, Power10 was supposed to be easier, I think, not harder, than Power9. And a lot of what we are seeing with Power10 now, including a truly new microarchitecture and up to 60 cores in a socket, was probably expected with Power11. This version of Power10, which we have covered extensively, is what we see in the most recent Power processor roadmaps:

A Power9+ chip implemented in 10 nanometers and a Power10+ chip implemented in 7 nanometers would have been extremely useful for IBM i shops, who don’t need more cores but rather faster cores and more cache memory and more main memory at a cheaper price and a lot more I/O coming into the system to balance this all out. So if the issues that IBM has had with its own Microelectronics foundry and with GlobalFoundries – and possibly will have with Samsung, its new foundry partner, as it does its first server-class chips – are disappointing in any way, it is that IBM has not been able to deliver the kind of Power Systems that would be truly useful to IBM i and AIX shops. And ditto for the partnership with Google, which seemed so promising back in 2013 when the OpenPower consortium was launched and when Google seemed poised to use a lot of Power9 chips to run its search engines and databases. (As far as we know, this effort has fizzled and Google doesn’t talk about it anymore.)

There’s no time quite like the present, and we remain hopeful that with Power10, Big Blue will create servers with very fast processors with a limited number of cores and lots of memory and I/O to make a screaming IBM i and AIX database engine. Hopefully this is in the cards, and hopefully the memory area network technology inside of Power10 will help IBM i and AIX shops create very powerful and very scalable clusters of machines that do interesting thing.

As for this lawsuit, maybe IBM will get some money to reinvest in the Systems division and maybe it will be tossed out of court before it goes to trial. Maybe the two parties will come to an arrangement. It is really hard to say. But we will be watching and keeping an eye out for how it might affect IBM i shops. Count on that.