Balancing Supply And Demand For Impending Big Power10 Iron
August 9, 2021 Timothy Prickett Morgan
It is always an exciting time when Big Blue is rolling out a new processor generation – well, with the exception of the Power6+, which IBM did not really talk about at all and tried to pass off as a Power6 until I figured that out. And I have to admit, IBM used the Power6+ architectural tweak and a refinement of the 65 nanometer chip manufacturing process (as opposed to an expected 45 nanometer process shrink) to still cram two whole processors into a dual chip module to radically expanding the throughput performance per socket, and it was impressive at the time.
Bringing chips to market is always tough. Always. There are always design issues and there are always manufacturing issues, and the fact that these ever-increasingly complex and ever-shrinking devices come to market at all is a bit of a miracle. A CPU socket is what we would have called an entire system (minus its main memory and storage) two decades ago. It’s amazing, and we owe the engineers of all chip designers and chip manufacturers some applause for the hard work they do and the engineering feats they pull off, always against the odds.
The Power10 chip has had its issues, of course, and I am glad for the sake of IBM’s Power Systems customers that the rollout is beginning. And it is doing so at the high end only because, I think, IBM is going for the richest vein of revenues and profits and is also covering the most demanding customers with what I presume is the limited supply of good Power10 chips it can put into the field in 2021. Hence, we are only going to see the Power10 chip in the Power E1080 machine this year, and from what I hear from the rumor mill, it is going to be a staged rollout as we saw with high-end, enterprise-class Power8 and Power9 systems in the past. (That’s the Power E870 and E880 based on Power8 and the Power E980 based on Power9.)
As I explained when we covered IBM’s lawsuit with former foundry partner GlobalFoundries back in June, Power9+ was supposed to be a dry run of 10 nanometer processes and a boosted NVLink interconnect and Power10 was supposed to be etched using a refined 10 nanometer process from GlobalFoundries and have a pair of 24-core chiplets to make a dual chip module. When GlobalFoundries spiked the 10 nanometer process, IBM had to spike the then-current Power10 design and shift to a 7 nanometer process that GlobalFoundries was working on. I don’t know what that design looked like – it could have been the same pair of 24-core dual-chip modules or it could have been the 15 fat core/30 skinny core Power10 chip that IBM revealed some of the specs of at the Hot Chips conference last year. Our guess is that the design changed considerably as the need for embedded machine learning became apparent and that the chip that is being manufactured by Samsung – it’s first server-class device – is different from the 7 nanometer chip design IBM had planned way back when with GlobalFoundries. To be precise, I think the addition of the mixed precision matrix math overlay for the Power10’s vector processors, which will be able to accelerate both AI and HPC workloads, is different, and to be even more precise, I have a strong hunch that IBM lowered core count to make each core to more work – and do a lot of work it was planning to offload to GPUs but now believes should be run natively on the Power10.
We have talked about the basic design of the Power10 chip here, drilled down into the Power10 architecture there, and talked about the possible designs of Power10 systems, including the Power E1080, at this link. I am not going to repeat all of that again. When I have all of the feeds and speeds and costs of the Power E1080 systems, I will tell you all about what IBM has delivered.
What I do want to let you know is that the scuttlebutt is that IBM will launch the initial Power E1080 machines sometime after the Labor Day holiday, and the rollout will be gradual. I am also hearing that it is more likely that the so-called scale out variants of the Power10 machines, which have one, two, or four sockets and which I expected towards the end of the first quarter – meaning February or March 2022 – are now looking like they will come out in the early part of the second quarter – meaning possibly April and May. That is, of course, subject to change.
It is helpful to remember that IBM is changing foundries, chip design tooling, microarchitecture, and manufacturing process at the same time. And branding, too, while I am thinking about it. This is a lot to change at the same time, and the only way to mitigate risk is to be extra careful. Considering the mission critical nature of IBM’s Power Systems business, taking the time to check and double check and triple check everything is absolutely the right thing to do, despite the kind of competitive impatience we have here at The Four Hundred.
The other thing is, IBM basically owns the market for big NUMA iron with RISC processors running a Unix or Linux operating system variant, and it is not under any competitive pressure to do anything other than make sure stuff works to keep its customers happy. I have heard there are over 10,000 customers running enterprise-class Power Systems based on Power7, Power7+, and Power8 machines, which probably means there is something around 13,000 to 14,000 customers once you add in Power9 machines. These customers represent the vast majority of the Power Systems revenue stream, but obviously a very small portion of the overall Power Systems base, which I estimate at 120,000 for IBM i shops and maybe another 10,000 to 20,000 or so for AIX shops. (There are big error bars on that AIX shop count. It could be a lot larger, or a lot smaller.)
The thing to remember is that with the initial Power10 rollout with the Power E1080 machine, IBM’s focus will be to move customers with those older Power7, Power7+, and Power8 machines to Power10 iron, and they will need a relatively few Power10 chips to accommodate their processing, memory, and I/O needs. So IBM will be able to take a relative few Power10 chips and cover many, many of these vintage Power Systems customers. The focus is not to sell large configurations of Power10 machines to a few Power9 customers; they can wait until Power10 chip volumes ramp and the supply chain eases.
We think there is a very good chance, in fact, that IBM will take a dual-pronged approach and push Power10 for those who really need it and Power9 for those who really do not. The issue comes down to how many Power9 chips IBM got in the barn before it sued GlobalFoundries in June of this year. Hopefully, a big number so it has optionality depending on how the chip manufacturing yield curve is going at Samsung and how that curve is shaped over the next year or so. I think yields are crap for all 7 nanometer processes, and I have no reason to believe that the first time out making a server chip is a cakewalk for Samsung. But I do believe that IBM and Samsung are working together and have figured it out, just as IBM and GlobalFoundries did to get 14 nanometer Power9 chips into the field.
I would not be surprised, in the long run, to see IBM dual-source design and manufacturing of its Power and z CPUs at Samsung and Intel Foundry Services, the new foundry operation that the world’s largest semiconductor maker set up earlier this year and which I wrote up in detail over at The Next Platform in March if you want to know more about that.