What The IBM Chip Biz Selloff Means To IBM i Shops
October 27, 2014 Timothy Prickett Morgan
IBM finally went and did it and not only sold off its X86 server line so it could focus on Power Systems and System z mainframes, but it went the next logical step and got itself out of the chip making business. Both decisions were the inevitable consequences of deals long since done, and their consequences will be felt in the decade hence.
In the case of the System x division, IBM’s fate was sealed at the moment that it sold off its PC division to Lenovo Group nearly a decade ago. Big Blue had very sound financial reasons for getting out of the PC business, where it had also ran far behind Hewlett-Packard and Dell. But the moment IBM sold off that PC unit for a paltry $1.25 billion, it lost a lot of its leverage with chip maker Intel. The fates of Advanced Micro Devices and IBM both might have changed dramatically for the better if IBM had instead bought AMD and decided to go head-to-head with Intel with its own PCs and X86 servers, and thereby kept its chip making fabs warmer. That leverage with Intel is what has given HP and Dell a leg up on IBM for the past decade in the X86 server business, as HP is just about to find out as it spins off its PC and printer group as a separate company. This is all good news for Dell, presumably, since it is holding onto its PC business, at least for now.
The fate of the Power chip business was put into jeopardy the minute that Apple ported its Mac PCs from PowerPC chips made by IBM and Motorola to Core and Xeon chips from Intel. IBM kept the volumes high in its East Fishkill, New York, wafer baker with the Cell variant of the PowerPC chip, which was used by Toshiba in consumer electronics as well as in the “Roadrunner” massively parallel supercomputer at Sandia National Labs, the first petaflopping machine and one, ironically enough, that mixed AMD Opteron and IBM Cell processors to create a massively accelerated hybrid supercomputer. Variants of the Cell and Power4 processors were used by Microsoft, Nintendo, and Sony for their game consoles, and the volumes of those machines kept the East Fishkill foundry warm for many years. But IBM lost that business to AMD a few years back–see, maybe you should have bought AMD instead of so many shares.
And since that time, the ARM processor, widely adopted in in smartphones, tablets, and all kinds of consumer and embedded devices, has captured the market share and mind share of companies that might have otherwise adopted a variant of the very well-engineered Power processors. Given that Apple was already on Power and was an original founding member of the PowerPC Alliance along with IBM and Motorola, it is a crime against engineering that Apple did not work with IBM to make its own variant of the PowerPC chip for the iPhone and then the iPad and that they could not come up with appropriate chips at affordable prices with merchant foundries to keep the PowerPC growing and vibrant.
The ARM architecture is a chip technology that can be licensed and modified and uses tools that are widely adopted by the few big merchant chip foundries, the most important ones being Taiwan Semiconductor Manufacturing Corp and GlobalFoundries. IBM’s Power architecture is tied very closely to its chip development tools and its own copper and silicon-on insulator processes. You can’t just take a set of Power8 masks and hand them over to TSMC or GlobalFoundries to start baking chips. It doesn’t work like that.
It is unclear how IBM made the decisions that it did. Perhaps it believed it could maintain the game console and embedded PowerPC business longer. Perhaps it thought the Unix market would not implode as it did. No one saw the Great Recession coming–well, honestly, we all must have suspected the punch bowl would be removed from the party at some point–or at least pretended they didn’t see it. But the data was always there: Intel was being aggressive, ARM was being aggressive, and Power, like MIPS and Sparc and Itanium, would be caught in the crossfire. But given IBM’s brilliance in chip research and development and its deep pockets and its steady System z mainframe and growing Power Systems business, IBM must have thought it could skate by. But IBM really needed to take the ARM approach and open up Power a long time ago. Kudos to Big Blue for creating the OpenPower Foundation and doing it, but if this had happened many years ago, then AMD might have licensed Power instead of ARM and GlobalFoundries might be already making Power processors at a time when IBM could not afford it.
Having said all of that, and yes it is frustrating for all of us as well as IBM, the fact that IBM is paying GlobalFoundries $1.5 billion to take over its foundries and to become its sole supplier of Power and mainframe processors for the next decade is better than the alternative, which is that IBM shuts down the fabs entirely or never progresses beyond the 22 nanometer node in process technologies. That is precisely where the Itanium chip is today: stuck at 32 nanometers with everyone already knowing it is dead but neither HP nor Intel ever admitting it.
As it was explained to me last week by John Kelly, director of IBM Research, the two are committed to delivering chips based on 14 nanometer and then 10 nanometer processes. The 14 nanometer node will use a mix of IBM and GlobalFoundries technologies as the Power and z chips are migrated toward the bulk CMOS processes used by the merchant wafer baker (which used to be AMD’s foundry operations, you will remember). The 10 nanometer node will be based on GloFo processes entirely. So, that 14 nanometer node is an interim step, like IBM i 6.1.1.
IBM has contended for years that getting below 10 nanometers would be very difficult, and in a presentation a few years ago, Kelly explained that eventually each node would require an entirely different process, not just a tweak on an existing process. It cost IBM about $2.5 billion to get its East Fishkill factor up to speed with 22 nanometer technology, and each successive node costs a little more. With IBM so hell-bent on buying back shares, it simply cannot make the necessary investments to stay in this game. In fact, if IBM had not already allocated the money and if it was on bulk silicon, it would have ditched for a merchant fab many years ago. The inertia in this process, which is almost geologic, makes change very difficult. Such transitions will take many years.
If this were a normal Power roadmap, I would say that Power8+ would be a shrink to 14 nanometers and Power9 might be a move to 10 nanometers. But that might have been optimistic. Power5 was 130 nanometers, followed by a shrink to 90 nanometer with Power5+. The Power6 chip was etched in 65 nanometers, with yet another shrink, but IBM didn’t make the 45 nanometer shrink with Power6+, which was actually not publicly launched as Power6 at the time even though the design was tweaked a bit. Power7 came out on 45 nanometer processes, and three years later Power7+ came out with 32 nanometer processes. Power8 came in at 22 nanometers, and I happen to believe that IBM is having trouble getting high yields on the process because it has very low volumes making chips. This is a volume business not just because you need volumes to make revenues and profits, but because you need volumes to tune processes and improve them. All of the costs for this ramp to 22 nanometers were being born, more or less, by the Power Systems and System z divisions now. So not only does it take more time, but it eats into profits.
See how this is all connected?
If this were a normal roadmap, the Power8+ would be a twelve-core chip that is a lot smaller and a lot faster than the Power8, and Power9 might have another 30 percent to 50 percent more cores. Call it 50 percent more to keep parity with the current top-end “Haswell” Xeon E5-2600 v3 processors, which have a maximum of 18 cores. Well, it is pretty clear that Power8+ will need more cores to compete, so IBM will have to do a chip design and a process shrink at the same time. That worked for Power5+, didn’t work for Power4+ or Power6+, and was of modest use with Power7+, and in each case the redesigns for the “plus” chips were not radical.
The funny thing is, IBM i shops don’t need more cores. They need faster cores and lower prices, and these are the two things that IBM is not interesting in giving them. Actually, to be fair, customers would take more cores at lower speeds if the prices were right on the hardware and the software. The IBM i operating system should not be priced by the core, but by the socket, freeing customers and IBM alike to get as much oomph in that socket as possible and use it as often as possible. But IBM’s model is to treat a core and an operating system as a dearly expensive thing rather than a commodity, and you can hardly blame IBM given how much money goes into making a Power chip and how much more it needs to extract out of customers.
IBM has faced a lot of problems with the Power architecture in the past decade, including the fact that its endian (the nature of the ordering of bits in the machine) was big instead of little like X86 chips. IBM has fixed this with Power8, and it has opened up the architecture with OpenPower so others can license and innovate. These are all good things. And the deal with GlobalFoundries to take over Power and System z chip production is also a good one–even at the high cost–because the alternatives would have been truly awful for IBM i, AIX, and Power-Linux shops. This was the best deal that Tom Rosamilia, the head of Systems and Technology Group, could do for CEO Ginny Rometty given all the variables in play. They inherited this problem–they didn’t create it.
And to his credit, Bill Zeitler, the long-time AS/400 Division general manager who went on to run Server Group and then STG when it was first founded, saw all of this and was very keen on expanding the use of Power chips beyond servers and PCs way back when. (Don’t think I have forgotten, Bill.) The one thing IBM did not do back in the mid-2000s when this plan was cooked up was truly open up the Power architecture. Of course, by then in early 2007, IBM had sold off its PC business two years earlier and the first Apple iPhone was due to launched within a few months. No one knew it would change everything, of course. Not even Steve Jobs. You can never know these things for sure, but you can hope and you can dream and you can keep moving.
And that is what the deal with GlobalFoundries is all about. The foundry will be getting 5,000 IBMers in the Vermont and New York facilities that will join the number two foundry in the world, which has a $12 billion plant, called Fab 8, near Albany, New York’s capital. Around 800 of those IBM Microelectronics employees are serious chip technologists that IBM very likely did not want to make part of the deal. Ditto for the thousands of patents relating to chips and materials science that IBM sold to GloFo as part of the deal. But without those people and patents, GloFo would not do the deal, and it is safe to say that these assets are why the chip maker did the deal and their absence in past negotiations was why it took so long to do the deal. IBM is keeping a small team of chip experts to work on systems, and it remains committed to its five-year, $3 billion effort to do advanced research on systems and chip designs. The new twist, and one that IBM probably did not want to do, either, is that the foundry will have access to the results of this research.
With IBM Microelectronics making only $1.4 billion in revenues in 2013 and losing $700 million on that, Big Blue could not keep doing this anymore. The future of the Power Systems platform itself was in jeopardy with that kind of bleeding. So IBM had to give away the semiconductor candy store to preserve its systems business, which drives a large portion of its revenues and even more of its profits (including software and services properly allocated).
We now know that there will be future Power and z chips for the next ten years, and GlobalFoundries will be the exclusive supplier of them. What we do not know is how aggressive IBM and GlobalFoundries will be with the chip roadmaps and whether they will try to catch up with Intel and keep pace with the ARM collective. Hopefully IBM was far enough along in the transition to Power8+ that we don’t have to wait more than 18 to 24 months for the next bump. Power7 came out in April 2010 and Power7+ hit in October 2012, some 30 months later, and Power8 debuted in April 2014, 18 months later. Those plus nodes take IBM more time for some reason, and they don’t always impress. In any event, an average of around two years between chip generations is too long for IBM to keep pace with Intel and the ARM collective, who are doing it in less time. IBM has a per-core performance advantage with the Power architecture, but it is not clear how long that will last if the time between generations doesn’t shorten.
What IBM needs to do now is be very specific about its Power roadmap with customers, rather than vague as it usually is. When Oracle took over Sun Microsystems in January 2010, it put out a five-year roadmap for Sparc processors and systems, and while there has been some wiggle here and there, the fact remains that Oracle hit those targets and, more important, brought some rigor to the Sparc development effort and cut down on the gee-whiz and risky approaches that lead Sun to lose its place at the top of the Unix market in the wake of the dot-com bust. IBM has been a lot more conservative and no less innovative with its Power chips. What customers now want to know is what the future will look like so they can get comfortable and continue investing in Power Systems. And that is going to require some changes in the way the IBM i stack is pitched and packaged, too. IBM can be focused on Linux-on-Power and scale-out clusters with accelerated Power8 processors goosed by GPUs, flash, DSPs, FPGAs or what have you, but it had better make it relevant to the 150,000 IBM i customers or the 100,000 or so AIX customers or the point is moot. Getting new customers is great, but not if you lose existing ones.
The one big concern I have now is that Power chips also get more expensive than they were when made by IBM itself. There is a lot of talk that the low volumes IBM has for its Power and mainframe chips means they will be more expensive when made by GloFo, and therefore adversely affect the competitive position of Power Systems and System z mainframes. As for the mainframe, IBM charges tens to hundreds of thousands of dollars to activate a core on these machines, and makes millions of dollars per year in systems software from System z mainframes. So I am not worried there. But I am a little concerned about how future Power processors will be priced and how IBM will make them competitive. Hopefully, the company will enlighten us all soon.