IBM’s Plan For Etching Power10 And Later Chips
January 7, 2019 Timothy Prickett Morgan
Last summer, GlobalFoundries, the chip making conglomerate comprised of the foundry businesses of AMD and IBM plus Chartered Semiconductor Manufacturing, put the kibosh on its planned aggressive ramp of 7 nanometer chip making technologies. AMD and IBM, who both depended on GlobalFoundries for their server chip manufacturing, obviously knew well before this announcement that GlobalFoundries was going to be halting development and production ramp for 7 nanometers, so they were not left in as much of a lurch as it might seem.
Lucky for both companies, there is more than one foundry that was trying to stay on the bleeding edge, but that doesn’t mean that the pickings are not getting slim out there in Fab Land. AMD certainly cannot go to Intel and ask it make its chips, and for two good reasons. First, there is no way in hell that Intel would do this at any price that made sense for a competitive AMD to be peddling chips against Intel, and second, Intel does not have its own act together on either 10 nanometer or 7 nanometer development and manufacturing. Intel’s own 10 nanometer chips are years late coming to market, and this is leaving the door open in the server racket for anyone who is partnered with Taiwan Semiconductor Manufacturing Corp, which has cutting edge 12 nanometer chips in production (the Nvidia “Volta” V100 GPU coprocessors were first there) and is well on the way to shipping 7 nanometer AMD “Rome” Epyc and Arm chips (probably the “Triton” ThunderX3 from Marvell/Cavium will be first) in 2019.
There was a fourth choice for an advanced foundry, and this is a very good thing for Big Blue. As it turns out, this particular foundry is the largest semiconductor manufacturer in the world, surpassing Intel back in 2017 as memory and flash prices went through the roof, and it has also been a partner in the OpenPower Consortium since February 2014, giving it insight into what IBM is doing with the Power platform. Perhaps most importantly, this company was also a partner with IBM and AMD in developing technologies and techniques for etching ever-smaller chips, including fundamental research at the 14 nanometer, 10 nanometer, 7 nanometer, and 5 nanometer nodes. The company is, of course, Samsung Electronics, the chip making arm of the $225 billion Korean conglomerate known as Samsung Group.
Bob Picciano, senior vice president in charge of IBM’s Cognitive Systems division, confirmed the choice of Samsung as IBM’s future processor foundry with The Four Hundred, and provided some insight into what IBM will be doing in the coming years.
The original plan for Power10 was to have it etched using 10 nanometer processes from GlobalFoundries and we did not expect to see a shrink to 7 nanometers until the Power11 generation of chips. The timing on these chips was always a bit vague because the technologies that need to be developed were somewhat vague, as is always the case with chip manufacturing, but as the transistor geometries are getting smaller and smaller, the challenges are getting steeper and steeper. So it is getting harder and harder to not only stay on the Moore’s Law pace of shrinking transistors every 18 months (the pace for decades) to 24 months (the pace for the past few years), but harder to stay on the CMOS chip making road for the long haul. But, we do not have any relatively cheaper alternative to CMOS, so we are going to have to make do for the next decade or so, and that will mean getting clever about compute and memory and storage packaging and interconnects. But that is a story for another day.
The IBM contract with GlobalFoundries was set to end in 2024 with the Power11 chips, and the two companies are still working together to make the current 14 nanometer Power9 chips and are collaborating on the future Power9’ (that is a prime sign) chip that will most likely come out later this year with some microarchitecture, memory, and I/O tweaks based on a refined 14 nanometer process from GlobalFoundries. At some point, if GlobalFoundries restarts its 7 nanometer efforts, there is a slim possibility that IBM could dual source Power10 chips, but this seems highly unlikely given how quickly GlobalFoundries ditched its 7 nanometer efforts, which including traditional water immersion lithography using visible light as well as more advanced extreme ultraviolet (EUV) lithography, which everyone – meaning TSMC, Intel, and Samsung – agree is the future of driving down the cost of making transistors in 7 nanometer and smaller geometries more affordable.
The good news for IBM i shops is that GlobalFoundries is doubling down on its 14 nanometer chip making at Fab 1 in Dresden, Germany, Fab 8 in Malta, New York, and Fab 10 in East Fishkill, New York. So Big Blue will be able to kick out Power9 processors for the foreseeable future as needed, and with Power10 probably not coming until late 2020 or early 2021 at this point. (The plan for Power10 was more like late 2019 to early 2020 using 10 nanometer processes, and doubling up the core count to 48 per socket.) For many IBM i shops, who have two cores, not processors, running their workloads, the current Power8 chip provides all of the computing and memory headroom they need. All IBM would have to do to meet their increasing workload needs would be cut the price of the cores in half every three years and the typical shop would see a 30 percent price/performance increase on average annually and that would give them 33 percent performance each year and IBM would not have to change the Power S914 or Power S924 systems at all. To be even more precise, a 48-core machine would do for such a shop with that very big growth in compute capacity until somewhere around the end of 2035, and it would still do the job for an IBM i shop with two cores running the platform until around 2033. That would be around when the Power14 chip would be expected, and when the Power14 might cram hundreds of processors onto a socket – most likely 3D multi-chip module with memory and nonvolatile storage embedded in a cube that is cooled by liquid nitrogen or some exotic thing. A socket will be a computer that this point.
The good news for Samsung, which thought about getting into the processor business a few years ago, is that it will be able to use IBM’s processors as a way of ramping its technologies. Under the current plan, the Power10 chips will be implemented in 7 nanometer processes and so will be the future System z16 mainframe processors; the System z15 mainframe motors will be etched using that same refined 14 nanometer process that is being used to create the Power9 chips. Beyond that, we expect refined 7 nanometer processes for successive Power and System z generations, and eventually a shift to 5 nanometer processors, with subsequent refinements. After that, we cannot see any further than the best minds in the chip industry, but we know one thing. IBM is in no worse shape than the rest of the processor designers who no longer own fabs and the one processor designer – Intel – that still owns them.
This is just a tough business, and it is getting tougher. And for the Power Systems business to do well means keeping up with Intel and using every advantage to try to steal away some of that precious datacenter profit away from Intel.
The Power9 chips are very competitive, technically, with anything Intel or AMD or the Arm collective can put into the field, but it is still a more difficult and risky choice. AMD’s Epyc processors are the easiest alternative to the Intel Xeon, and it is arguable if the Power or Arm processors are better than each other as the next option. But for IBM to stay in the processor design and system building game, Power chips have to keep evolving and they have to keep gaining traction in machine learning, supercomputing, data analytics, traditional database processing, and in-memory and stream processing. It has to become a $4 billion to $5 billion server business for IBM and its OpenPower partners, as the Power chips drove the former AS/400 and RS/6000 businesses back in the late 1990s and early 2000s. Big Blue needs to make money from this because it is not a chip charity. It is a public company with shareholders that are as demanding as its customers.
IBM has all of the right ideas, and only a fool would count it out. But a resurgent AMD and a feisty Intel will still be tough to beat because of all of the preferences and prejudices out there. Watch carefully what the HPC centers and hyperscalers of the world do – and don’t do. There is your leading indicator to the future for the IBM i platform.