Ibm's Power11 Processor Architecture
Key topics
The mystery surrounding IBM's Power11 processor architecture has sparked a lively debate about who exactly is buying and using these high-end systems. While some commenters, like vb-8448, point out that there are still plenty of IBM-I and AIX installations around, others, like Someone, counter that the real question is how many new systems are being sold. A humorous exchange between CursedSilicon, gosub100, and others reveals that some people believe IBM AS/400s might be lurking in unexpected places, like Costco's back offices, although this is disputed by experts like dardeaup and gt0, who clarify that IBM's POWER lines are distinct from their mainframe line. Ultimately, robotnikman sheds some light on the matter, explaining that most stores likely use local servers to handle transactions, which are then batched and sent to a central server.
Snapshot generated from the HN discussion
Discussion Activity
Very active discussionFirst comment
54m
Peak period
37
0-3h
Avg / period
7.7
Based on 69 loaded comments
Key moments
- 01Story posted
Aug 25, 2025 at 1:43 PM EDT
4 months ago
Step 01 - 02First comment
Aug 25, 2025 at 2:37 PM EDT
54m after posting
Step 02 - 03Peak activity
37 comments in 0-3h
Hottest window of the conversation
Step 03 - 04Latest activity
Aug 27, 2025 at 7:08 AM EDT
4 months ago
Step 04
Generating AI Summary...
Analyzing up to 500 comments to identify key contributors and discussion patterns
Want the full context?
Jump to the original sources
Read the primary article or dive into the live Hacker News thread when you're ready.
They simply have to have some farms in government running this to make it make sense.
Also, how does “plenty” compare to the millions (50 million or so, it seems from a quick search) x64s that Intel sells per year? Do they even sell 1% of that?
On the high-end? Banks, airports, hospitals, research labs. There's a lot of places that need the kind of fault tolerance that specifically IBM POWER systems provide
EDIT: Okay, IBM POWER "systems". They've been described as mainframes to me so I went with that terminology
I think a more realistic case is the visa and MasterCard credit card networks that have almost 0% downtime.
One of my Summer intern jobs was to run backups every few days on a AS/400 system.
That "seat" was the old one that was yet to be collected.
Back then their model lineup ranged from the small size like you mention to approx 3 or 4 refrigerator size at the high end.
When I did some system work out at Costco in the 90's they had 4 of the largest models connected together in one system image (sysplex I think).
The smallest iSeries I've seen was essentially an "oversize tower PC" form factor.
Payments processed by the payment terminals handle the authorization of the payments separately from any store servers, usually through a service such as Connected Payments.
For AS/400 (IBM i), they're POWER, but come in pretty small models like IBM Power S1012, it's available as "deskside", i.e. a big tower.
For many years, the 64-bit extension of the original S/360/370/390 architecture was emulated in the software layer via the static binary translation – just like the i Series AS/400 have been doing since the inception, and there was no native S/360 implementation in silicon for a fairly long time.
If my understanding is correct, with TELUM processors, IBM has gone back to implementing the ISA in silicon, although the available details on TELUM are scarce.
The eCLIPz project, for the POWER6 & Z10[0].
"The z10 processor was co-developed with and shares many design traits with the POWER6 processor, such as fabrication technology, logic design, execution unit, floating-point units, bus technology (GX bus) and pipeline design style, i.e., a high frequency, low latency, deep (14 stages in the z10), in-order pipeline.
However, the processors are quite dissimilar in other respects, such as cache hierarchy and coherency, SMP topology and protocol, and chip organization."[1]
[0] https://news.ycombinator.com/item?id=18494225
[1] https://en.wikipedia.org/wiki/IBM_z10
IBM z15 mainframes had z15 chip now they have Tellum. z chips are their own line. z14, z13, ...
https://en.wikipedia.org/wiki/IBM_i#TIMI
To illustrate why the AS/400 had its market niche, the Cali Cartel had a very successful installation of an AS/400. Apparently they used it to do both "business analytics" and back-office tasks.
https://www.vice.com/en/article/the-cartel-supercomputer-of-...
The software translation layer has been a feature of the platform dating back to the System/38 days, and was specifically intended to allow the CPU architecture to change without breaking software compatibility.
IBM i also has the "PASE" layer, which is a binary compatibility layer for AIX. Those applications do not use the translation layer.
I suppose the military should buy a significant amount, because they require 100% US-based development and manufacture.
So is it science/defense that makes up their user base? I have an acquaintance who works at NASA mention that IBM stuff was really good for processing large amounts of data fast. Is that their primary advantage over other architectures?
edit: thought some ibm mainframes used power in helper roles.
For the embedded market, NXP makes a bunch of QorIQ chips based around the POWER ISA - mostly for telecom products. These are actually reasonably common in certain devices, but not really what you'd want in a desktop.
On that note, out of curiosity, how do these compare with x86 and ARM chips from that era?
I used to own one of these systems before I sold it to my dad, and I'm eagerly awaiting a next generation offering, if it ever comes out.
I really want IBM's Power ISA to have more mindshare, as I think it's a cleaner design and more straightforward than the likes of ARM or RISC-V. Unfortunately, IBM is more concerned with wringing out every last dollar from their contracts instead of pushing the ISA out as much as they can. OpenPOWER was supposed to be this initiative, but it has seemingly fizzled out to a large extent, such that NVidia and Google no longer are Platinum members and hardly anyone talks about it anymore.
[0] https://www.servethehome.com/ibm-power11-launched-with-up-to...
[1] https://www.redbooks.ibm.com/redpieces/pdfs/sg248587.pdf
Though based on what I've read so far, it all feels like vaporware.
But one thing I never knew about before was covered in just a few seconds in the video: 40+ layer motherboard PCBs. Holy cow, it was like a half inch thick! It makes sense now that I think about it considering how many components are slotted into it, but for some reason it never dawned on me that you could have a PCB that was so huge vertically.
In addition to the technical parameters mentioned in the article (an IBM Red Book would certainly provide more):
From what I know from conversations, SAP users are a large customers base for Power systems. I think the size of the possible main memory might be one reason, persistent main memory [1] could be another, in my opinion. But this feature is not new, it was available at least with Power 9. The systems are available with up to 64 TB memory (I don't know if all this memory would by available for one process, but from what I have read at least 32 TB is) and that seems to mix well with SAP HANA.
Another reason is licensing. LPAR partition is considered hard partitioning for Oracle databases, which makes licensing much easier and, in many configurations, also more cost-effective. [2]
Then there are several RAS features. Think of: if one core of a processor in a x86 server dies, your server dies. This is not the case in the IBM Power environment. These days many customers are happy with horizontal redundancy, but not all are equal.
[1] https://community.sap.com/t5/technology-blog-posts-by-member...
[2] https://oraclelicensingexperts.com/oracle-licensing-virtuali...
> if one core of a processor in a x86 server dies, your server dies. This is not the case in the IBM Power environment.
I believe that having a spare processor core is new to POWER11. EDIT: No. I was wrong.
Regarding the processor: yes the processor spare is new to the Power11 but features like "Processor Instruction Retry" and "Alternative processor recovery" are older. Here are documents describing the feature for old 8284-22A (S822 with Power8 CPU) [1] or even older 8204-E8A (System Power 550 with Power6 CPU) [2]
[1] https://www.redbooks.ibm.com/redpapers/pdfs/redp5098.pdf (topic 4.3.4 and 4.3.5) [2] https://www.redbooks.ibm.com/redpapers/pdfs/redp4404.pdf
https://wallaroo.ai/optimizing-ai-inference-and-governance-w...
A crazy friend of mine had a workstation with a Power10 main board in it running Redhat Linux. Other than the fact that it used what was essentially a rack server main board so had these 8 fans that blew through the chassis like a 747 on take off it was pretty neat.
It seems like these systems should be stupidly good at AI training.
What am I missing?
From my, admittedly uninformed, position, AI training requires big memory and enough processing units to keep the bandwidth to that memory fully utilized--things which Power should be very good at.
Really looking forward to x86 getting similar: Intel's MR-DIMM and industry standard MCR-DIMM are both sort of happening. Intel's Granite Rapids (Xeon 6900P series) for example takes either DDR5-6400 or MRDIMM-8800, a 38% increase. And which should also allow a capacity increase, up to 256GB dimms supposedly. But so far, seems like 96GB is as big as we get, and MCR-DIMMs are still work-in-progress. https://www.phoronix.com/review/intel-xeon6-mrdimm-ddr5 https://www.serversupply.com/MEMORY/PC5-70400/96GB/MICRON/MT...