"special Register Groups" Invaded Computer Dictionaries for Decades (2019)
Key topics
The fascinating tale of "special register groups" has resurfaced, sparking a lively discussion about the evolution of computer architecture. The original author, kens, weighed in, revealing that the Honeywell processor's "special register groups" employed a barrel processing technique, switching tasks and registers after every instruction. Commenters jumped into the fray, sharing examples of similar hardware multiprogramming concepts, such as the CDC 6600 and Tera/Cray's MTA, which boasted 128 register sets to handle 128 threads. As the conversation unfolded, it became clear that the term "special register groups" might be a bit of a misnomer, with some arguing that modern CPUs have plenty of specialized registers, while others pointed out that the terminology isn't commonly used in the industry.
Snapshot generated from the HN discussion
Discussion Activity
Active discussionFirst comment
18m
Peak period
18
0-3h
Avg / period
5.8
Based on 46 loaded comments
Key moments
- 01Story posted
Aug 26, 2025 at 2:57 PM EDT
4 months ago
Step 01 - 02First comment
Aug 26, 2025 at 3:15 PM EDT
18m after posting
Step 02 - 03Peak activity
18 comments in 0-3h
Hottest window of the conversation
Step 03 - 04Latest activity
Aug 28, 2025 at 3:16 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.
TI had some IC where the registers were in main memory, with a hardware register to point to them. Context switching didn't require a register save, just pointing the register pointer somewhere else. There were stack machines such as the B5500 where the top items on the stack were in hardware registers but the rest of the stack was in memory. Floating point in the early years of floating point co-processor chips had lots of rather special registers. Getting data in and out of them was sometimes quite slow, so it was desirable to get data into the math co-processor and work on it there without bringing it back into memory if at all possible.
Lots of variations on this theme. There's a lot going on with special-purpose registers today inside superscalar processors, but the programmer can't see most of it.
https://spectrum.ieee.org/the-inside-story-of-texas-instrume...
It doesn't seem too unreasonable to call the program counter, memory registers, etc. "special registers" since they have specific functions beyond simply storing data. Therefore I could imagine calling the aggregate of such registers in a CPU a "special register group". Perhaps those still using the term are following this line of thought.
So this sort of "quiz" is bogus because it's basically checking whether you've rote memorized some particular words. People who know how a CPU works are confused and likely don't pass, those who've memorized the words but lack all understanding pass. This is not a useful education.
My recollection is that even the Atmega 8-bit microcontrollers have tons of special register groups around timers and interrupts.
Given the relative scarcity of "general purpose" registers on x86 32-bit CPUs, you could actually argue those are the special purpose registers.
Not precisely. AVR, like most embedded architectures, has a bunch of I/O registers which control CPU and peripheral behavior - but those are distinct from the CPU's general-purpose registers. The I/O registers exist in the same address space as main memory, and can only be accessed as memory. You can write an instruction like "increment general-purpose register R3", but you can't use that same syntax for e.g. "increment the UART baud rate register"; you have to load a value from that register to a GPR, increment it there, and store it back to the I/O register.
AVR is a bit weird in that the CPU's general-purpose registers are also mapped to main memory and can be accessed as if they were memory - but that functionality is rarely used in practice.
Getting back to your original point, x86 does have special-purpose registers - lots of them, in fact. They're accessed using the privileged rdmsr/wrmsr instructions.
"special" is meaningless without a the presence of contrasting subjects that are "ordinary" or "regular".
E.g, "The human hands have two special finger groups".
One of my favorites is BLDC which stands for BrushLess Direct Current, which is used to describe motors which are driven by an alternating current at a variable frequency, not to be confused with VFD, which in this case stands for Variable Frequency Drive, and describes how BLDC actually functions, and is on rare occasions used interchangably, but is most often used with analog or open-loop controllers, while BLCD is used for VFDs that are digitally controlled. BLDCs are Brushless not Brush Less and, AC not DC. I like to think the acronym stands for Bitwise Logic Driven Current, which follows the acronym, correctly describes how it works, and differentiates it from non-BLDC VFDs. Also, while VFD can mean Variable Frequency Drive, it is also used to mean Vacuum Fluorescent Display, and for Lemony Snicket fans it can me a whole lot more. (https://snicket.fandom.com/wiki/V.F.D._%28disambiguation%29)
Uncommon jargon is even better, especially when it comes out of marketing departments and describes something ordinary. I have a rice cooker that has terms like "micom" and "fuzzy logic" on the labels. They describe microcontrollers and variables, respectively, which are found in all but the simplest electromechanical appliances.
It's normal for companies to trademark their names for common technology, like Nvidia and AMD trademarking their implementation of variable framerate as G-Sync and FreeSync, respectively, and some companies are more agressive with their trademarks than others, such as Intel trademarking their symetric multithreading as Hyper-Threading, but AMD just calling it symetric multithreading.
Cessna used to have a bunch of random trademarked jargon in their ads in the 70's, from trademarking flaps to back seats.
As far as "fuzzy logic", there's a whole lot of history behind how that ended up in your rice cooker. In the 1990s, fuzzy logic was a big academic research area that was supposed to revolutionize control systems. The basic idea was to extend Boolean logic to support in-between values, and there was a lot of mathematics behind this, with books, journals, conferences, and everything. Fuzzy logic didn't live up to the hype, and mostly disappeared. But fuzzy logic did end up being used in rice cookers, adjusting the temperature based on how cooked the rice was, instead of just being on or off.
[1] https://bitsavers.org/pdf/ibm/650/22-6270-1_RAM.pdf
[2] https://en.wikipedia.org/wiki/Fuzzy_logic
Actually it was a drum. No moving arms at all, but some tracks had multiple read/write heads to reduce access time.
Knuth's high school science fair project was to write an angular optimizing assembler for the IBM 650, called SOAP III.
"Records in the IBM Random Access Memory Unit are stored on the faces of magnetic disks."
"The stored data in the Random Access Memory Unit are read and written by access arms."
"The IBM 355 RAM units provide extemely large storage capacity for data... Up to four RAM units can be attached to the 650 to provide 24,000,000 digits of RAM storage."
The main memory on the other hand: "The 20,000 digits of storage, arranges as 2000 words of memory on the magnetic drum..."
I had looked up the history of fuzzy logic, when I originally saw the term. Everything I came across was from the mid 60's during the minicomputers era (e.g. the DEC PDP and IBM System/3). I just did a Google ngram search for it, and it does look like it peaked highest in the mid 90's. (https://books.google.com/ngrams/graph?content=fuzzy+logic&ye...) It was especially redundant then, because by that point 32-bit floating-point processors were already extremely common. It made sense during the 60's to have two-bit values in ladder logic control systems, because mini computers were prohibitively expensive, but by the 90's control systems were already using personal computer components to emulate old relay logic, while newer control systems were being written in languages that are still common today.
You just sent me down a rabbit hole of mid-90's buzzword ands jargon, and I found something especially great: neuro-fuzzy logic (https://link.springer.com/article/10.1007/s44196-024-00709-z)
There's also quantum logic (https://en.wikipedia.org/wiki/Quantum_logic) and someone wrote a paper relating it with fuzzy logic: (https://link.springer.com/chapter/10.1007/978-3-540-93802-6_...)
Once academia gets involved, you get fractal (https://xkcd.com/1095/) sub-niches of sub-niches, and it goes further off the rails than marketing jargon.
I swear there's a computing buzzword cycle of embedded systems (e.g. fuzzy logic, smart, IoT, etc.), artificial intelegence (e.g. genetic algorthims, machine vision, neural networks, etc.), and quantum (just quantum, it's never been practical enough to have sub-niches). Right now we're on the trailing end of AI and working our way to quantum. There's a tech company called C3 that has changed their name from the original C3 Energy to C3 IoT and now C3 AI. When they change it to C3 Quantum, we'll know were in the next phase of the buzzword cycle.
At the windings, all motors are driven by some alternating waveform. A classic "DC motor" has a mechanical commutator which turns DC into square wave AC, with the phase leading the motor so the motor turns. Classic AC motors are driven from sinusoidal waveforms. There's a whole theory of DC motors, and an elegant theory of AC motors that goes back to Tesla. Here's the motor family tree.[1]
Then came power MOSFETs. Today you can make pretty much whatever waveforms you want. It took a while for motor designers to learn how to exploit that properly, and for MOSFETs to get small, cheap, and heat-tolerant. Then drone and electric vehicle motors got really good, at the cost of needing a CPU to manage the motor.
[1] https://www.allaboutcircuits.com/textbook/alternating-curren...
For example, how many things does "link" mean? "Process"? "Type"? "Local"? It makes people (e.g., non-technical people) think that they understand what I mean when I talk about these things but sometimes they do and sometimes they don't. Sometimes we use it in a colloquial sense, but sometimes we'd like to use it in a strict technical sense. Sometimes we can invent a new, precise term like "hyperlink" or "codec" but as often as not it fails to gain traction ("hyperlink" is outdated).
That's one reason we get a lot of acronyms, too. They're too unconversational but they can at least signal we're talking about something specific and rigorous rather than loose.
I'm going to guess it's a Japanese rice cooker, because "micom" is a Japanese shortening of "microcomputer", which is what they've been calling microcontrollers. Using it for marketing most certainly dates from the time when computerised control was considered a desirable novel feature, much like when "solid state" was used in English.
Japan loves to give words back to English though, with cosplay, anime, emoticon, and now micom all originally being from English, then used in Japanese, and brought back to English from their Japanese form.
It's a thing that happens with linguistic contact, and the deeper learning is that it's odd to expect it not to happen to the same word twice between languages with long-standing contact on occasion, given how often it happens and for how many millennia the process has been active.
How do you produce a final "m" in Japanese without a following M, B, or P?
I'm surprised that the analogy to hyperthreading wasn't made here.
The first CPUs with SMT have appeared only during the nineties.
That Honeywell CPU is the first commercial CPU with FGMT (Fine-Grained Multithreading).
SMT can exist only in a CPU able to decode and execute multiple instructions simultaneously, which is why it has appeared much later.
While FGMT went out-of-fashion for CPUs, many GPUs have used and many still use FGMT, because it can provide better performance per die area than a more complex superscalar/SMT architecture.
https://www.bidspotter.com/en-gb/auction-catalogues/quaker-c...
I would estimate it was fine-grained.
No microprocessor in the CPU, just a PCB a foot square covered in discrete logic chips.
Memory PCB the same size, connected by short ribbon cables to the CPU, this bypassed the backplane. Fully upgraded it was 16K of 32-bit memory.
Each user could run more than one Basic program at a time, and also the same Basic code in memory could be run by more than one user, or even more than one time simultaneously by the same user. Simultaneous with the collection of live data from more than one chemical analyzer, which the firmware could handle most of the mainstream chemical needs without the Basic option. User-programming was for the exotic stuff that actually requires custom code.
But that made it completely programmable using only the documentation supplied, it was straightforward to accomplish things which PC software that eventually came along never has been able to do without settling for additional commercial offerings which are much more user-friendly (mouse!) but not as appropriate as my own code was.
This avoids wasting clock cycles in threads that are waiting for some event.
My favorite example is the "tongue map" - decades of schoolchildren have been taught that different parts of the tongue are responsible for tasting sweetness and saltiness and so on. This is contrary to everyone's experience, and it turns out to have been a mistranslation of some random foreign journal article on taste buds, but it's stuck around in the primary school curriculum because it's easy to make it into a "fill in the map" activity. As long as the kids can regurgitate what they're told, who cares if anything they're learning is true?
1. https://lowell.edu/percival-lowells-search-for-life-on-mars/
[1]: https://en.wikipedia.org/wiki/Martian_canals#Supposed_%22dis...
“Special register groups” invaded computer dictionaries for decades - https://news.ycombinator.com/item?id=21333245 - Oct 2019 (60 comments)
The term CPU has been introduced by IBM in 1954, in the manual of operation of IBM 704, and it was defined thus: “The central processing unit accomplishes all arithmetic and control functions”.
This definition was clearly meant to say that the CPU is the aggregate of 2 parts of the 5 computer parts listed by John von Neumann, i.e. the CPU is the "central arithmetical part" + "the central control part" from von Neumann's list. The other 3 parts were the main memory, the input peripherals and the output peripherals.
The IBM definition has remained valid until today, even if the package that contains the CPU also contains cache memories and possibly even a part or all of the DRAM memory, like in Apple CPUs or Intel Lunar Lake.
Because the CPU is the most important part of a CPU package, the whole package is referred as the CPU ("pars prō tōtō"), even when the package also contains some part of the memory.
I assume that the same way of using the "CPU" term has caused the confusion from the dictionary entry.
In IBM 704, the CPU filled an entire cabinet and the main memory was in a separate cabinet. In later computers, by the time of Honeywell 800, it became possible to locate both the CPU and the main memory in a single cabinet, which was referred to as the CPU cabinet, as the CPU was its most important part.
Whoever has copied that CPU definition, presumably from a manual of Honeywell 800, has confused the description of the content of the "CPU cabinet" of that particular computer with the definition of the term "CPU" as the name of a block used in the description of computer architectures.