To Comfort The Afflicted And Afflict The Comfortable

July 14, 2025 Timothy Prickett Morgan

We are still assembling the information on the new Power11 processor and the Power Systems machines that use them and working on our analysis before they start shipping on July 25. We will do the most thorough job that we can under the circumstances, as always, but we concede that pricing information is very hard to come by on Power10 systems and we suspect it will be the same with Power11 systems, too.

But we will remind IBM that any new generation of machines always has improved bang for the buck, and given the modest improvements in performance moving from Power10 to Power11 and the desire by IBM to move Power9 shops to Power11 – almost all of the comparisons it is making are for that bigger jump, which has its complications – then it has got to cut prices to deliver a reasonable price/performance improvement. A lot will depend on the software tiers that people move from and to, whether they go with Power10 or Power11 iron, and how they adopt subscription pricing or not.

There are a lot of different things to consider, and we will walk through them. Feel free to send us the data you have and your thoughts about Power11 and what to do with it – and when – as you develop them. Together, we are a giant distributed human inference engine and I like that about us.

But before we get into all of that, I thought it was important to frame the discussion, and talk a bit about the Power10 and Power11 chips we got and the ones that we might have otherwise gotten if IBM didn’t have so many issues with the transition from 22 nanometer chip manufacturing with the Power8 processors from 2014 down to 7 nanometer processes used for both Power10 and Power11 in September 2021 and in July 2025, respectively.

There is an old adage, from 1902 by Chicago humorist and journalist Finley Peter Dunne, who was writing in the persona of an Irish bartender named Dooley when he quipped that “the job of the newspaper is to comfort the afflicted and afflict the comfortable.”

I didn’t study journalism so much as being thrown into it to try to make a living in New York City, and my mentor, the late and definitely great Hesh Wiener of Brooklyn by way of Ellenville, New York up in the Catskills a bit north of where I grew up, drilled this idea into my head over cigarettes and scotch on Friday editing nights when we put The Four Hundred and other newsletters he published together before they were sent off to the printer. Yes, we predate the commercial Internet by a decade and we used to be a subscription newsletter printed on paper.

Afflict the comfortable and comfort the afflicted. Hesh reversed the order of operations when he explained the job. And I have come to realize that it is not really me that changes when I do either of these tasks, but the subjects of my analysis. We got to know IBM as a reformed serial monopolist, twice convicted, and also watched its near-death experience in the early 1990s, and did our part with constructive and hopeful criticism as it went through its difficult times. We watched as Sun Microsystems, Microsoft, Intel, Google, Amazon, Meta Platforms, Nvidia, and others (perhaps someday OpenAI) have gone through their rises to prominence and power.

Nvidia is a significant case in point here in 2005. We watched Nvidia enter the datacenter, unintentionally at first, and become a force to be reckoned with as it sells a new computing paradigm we call GenAI, driving huge revenues and profits – ones that dwarf levels attained by Big Blue at its height, as hard as that might be to believe. And we have watched as Nvidia starts to inevitably change, as all companies in its position of influence and power inevitably must as the Nvidia top brass fulfill their fiduciary responsibilities to increase sales and profits.

We admire Nvidia a great deal, but we will call it as we see it if Nvidia steps out of bounds with its customers, partners, or adversaries, just as we always did with IBM.

These days we are more apt to comfort IBM and its customers, particularly the IBM i base, for which we are an avowed advocate. This is particularly true as the Power11 comes to market. Arguably, the Power11 chip does not represent much of an architectural or manufacturing leap, and we have said that to one way of thinking about it, this is really just a deep bin sort on a refined 7 nanometer process from Samsung that we might call Power10+ or perhaps more accurately a Power10’ chip. That apostrophe is short for “prime” which signifies even less of an architectural or manufacturing process change than a “plus” on a Power chip generation’s tweaked variants.

We went over the many Power roadmap changes that IBM has endured back in June 2021 in IBM Versus GlobalFoundries: A Lawsuit Instead Of The Power Chips Planned. But it bears some repeating on the main points.

Here is the Power roadmap as the Power9 chip was being delivered in late 2017:

At that time, IBM was going to shrink to 10 nanometer processes and double up the core counts on the chip to 48 cores. The cadence on Power chip announcements was slated for three years, with an interim Power8’ (with Nvidia NVLink ports added) delivered and a Power9’ (with faster memory subsystems prototyping what was to be delivered in Power10 added) on the roadmap but not shown here.

GlobalFoundries had issues with its 10 nanometer processes, and eventually convinced IBM to switch to 7 nanometer extreme ultraviolet (EUV) processors to deliver a much better Power10 chip to market, but later than expected. And so IBM rearchitected the Power10 to be a totally different chip, one aimed more at IBM i and AIX customers and less at trying to take on the X86 processors from Intel and AMD on their home turf in the datacenter. We did not expect to see 7 nanometer processes used until a Power10+ chip or maybe a Power11 chip.

That 7 nanometer shrink at GlobalFoundries was supposed to deliver up to 60 SMT4 cores or up to 30 SMT8 cores in a single socket, we think using dual-chip modules but perhaps not. Take a look:

GlobalFoundries pulled the plug on 7 nanometer EUV, and IBM had to jump to Samsung as its foundry, and Samsung had a 7 nanometer process under development that would be good for server CPUs, with IBM as its flagship customer. So instead of a Power10 with 24 fat SMT8 cores and 48 skinny SMT4 cores, we ended up with a totally new Power10 with a revamped instruction set, matrix math units in the cores, and a commercialized differential DIMM memory subsystem based on OpenCAPI memory ports. As far as we know, there never was a version of Power10 with SMT4 cores offered to customers. And the yields on the original 7 nanometer processes from Samsung were never high enough to have all 16 cores on the Power10 dies activated.

Were it not for the difficulties with GlobalFoundries, IBM may have continued down the path of trying to compete more directly with X86 CPU designs rather than veering onto a path that better suits IBM i and AIX customers and improving the chips in ways that are valuable to them.

In a sense, the Power10 we got was what IBM might have been planning with Power11, and that makes the Power11 we will get in a few weeks a little bit like a Power11’ chip. If there had been a process shrink to 5 nanometers, which would have allowed the clock speeds rise a bit more, but at tremendous cost to IBM that might have wiped out a lot of its profits from Power Systems hardware sales, this could have legitimately been called the Power11+ chip. Such a chip might have been better for customers, but bad for the Power Systems division, ironically. Which in the long run, would be bad for customers.

All of this said, we think the bar is set pretty high for a lot of innovation and a pretty big process shrink for Power12, which will employ a more modern chiplet architecture than Big Blue has delivered to date. But, once again, driving the bleeding edge of process technology, say at 2 nanometers or smaller, may not be worth it. So don’t be surprised if IBM doesn’t have Samsung push process as aggressively as it otherwise might.

With that said, we can now begin to drill down into the Power11 chip and its systems.

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search