Julia Evans
everyone keeps asking me why x86 uses an 8-bit byte, but I'm struggling to find an explanation that makes sense to me. Can any of you help?
what I've found so far
- it looks like x86 evolved from the intel 8008 (from 1972), which was an 8-bit CPU
- the 8008 came after the 4004, which was 4-bit
some questions I have:
- was the reason to build an 8-bit CPU to increase the size of the instruction set? or something else?
- did x86 really directly evolve from the intel 8008?
would love any links!
Robin
@b0rk @palin
Not an answer, but a conjecture: 4 bytes can store one digit in BCD format, but 3 can't. And there was a time (probably still is, cos banks) that BCD was a thing.
clacke
Julia Evans
@palin I think we still have the distinction between bytes and words, but nobody really talks about nibbles anymore
(I actually don't understand why people ever talked about nibbles in the first place -- I'd be curious to understand that too)
Palin
@b0rk well back in the days we had a distinction between bytes (8 bits), nibbles (4 bits) and words, which were different based on how many bits the cpu processed in a single instruction (8 bits had 8 bit words, 16 bits 16, and so on).
Never occurred to me there were non 8-bit bytes.
Peter Drake, he/him, LFHCfS 🔥
@b0rk I don't have any useful links other than the one you probably looked at first:
https://en.wikipedia.org/wiki/Byte
I remember (but can't find) a list of alternate historical sizes. My favorite was the 13-bit "baker's byte".
Just-so stories I can't back up with sources:
A power of two is convenient as a memory size. It takes three bits to address the bits in an 8-bit byte, but it would take four (with the last one partially wasted) to address the bits in a 10-bit byte.
If you want a character set that includes upper- and lower-case English letters, digits, and some punctuation marks, you're going to need at least 7 bits. I believe the 8th bit was originally used for error detection.
Julia Evans
or maybe the reason was that the 8008 was a popular microprocessor, and it happened to use an 8-bit byte so it became the foundation for all of intel’s future microprocessors, but in theory they could have also built a microprocessor with a 10-bit byte instead and that would have been fine too?
clacke
Eʟʟ
@b0rk This is one of my favorite parts of this whole thing! I don't think there's a clear technical reason, given that we had 6-bit computers as well - even numbers are nice because you can split them into even nibbles, and powers of two feel familiar, but since bits aren't individually addressable, I don't think there are any material reasons to prefer powers of two.
I don't have any links on this, but one pressure keeping bytes from 10 bits is wasted space - how often are you going to use those top two bits? 256 is a pretty good balance, you have room for Latin text with some accents
Steve Bellovin
@b0rk @oclsc You might want to read this bio of Fred Brooks (https://en.wikipedia.org/wiki/Fred_Brooks)—he said that the decision to use 8-bit bytes on the IBM S/360, to permit use of lower-case letters, was his most important. (Aside: Brooks, who died last November, was one of my mentors. Saturday, I was in North Carolina for a "professional memorial" to him, about which I may post more later. For more on my relationship with him, see https://www.cs.columbia.edu/~smb/blog/2022-11/2022-11-18.html).
clacke
Diane 🕵
@b0rk early machines had other wacky word sizes. Like 10, 12, or 36
https://en.m.wikipedia.org/wiki/Word_(computer_architecture)
The page also suggests the 8 bit byte became popular after ASCII was standardized
Julia Evans
so far the reasoning I'm getting for the 8-bit byte seems to be:
1. you want your byte size to be a power of 2. This is EXTREMELY believable to me, but I don't understand why exactly you want this, my understanding of CPU design is very bad. Maybe it's because of busses? (what's a bus?)
2. 4 bits is too small, you can't fit a character into 4 bits
3. you also don't want your bytes to be too big, and 8 bits was working well, so the byte size never got bigger after the move from 4 to 8
clacke
clacke
Dan Lyke
@b0rk Place where this got super expensive: The Apple ][ used 7 bit words in its hi-res framebuffer (even one color, odd another, high bit chose palette for those pixels). The hoops we jumped through to try to divide by 7, or build sprite animations around 7 or 14 pixel cycles, were legendary. Pretty much everybody did it as a lookup table, and that's 280 bytes.
2/fin
Dan Lyke
@b0rk I thought I saw this addressed earlier, and my head is fuzzy from allergies or a cold, so apologies if this reply is redundant, but something like 10 bits means that to, say, find the 13th bit you actually have to divide by 10. Dividing by 8 you can do with an AND (for the remainder) and an SHR for the divide.
And in hardware, those operations map to transistors and circuit complexity.
1/n
Julia Evans
my main question about this 8-bit byte thing right now is -- everyone says that's it's good to have a byte size that's a power of 2. That sounds very plausible, but why exactly is it important to use a power of 2?
(obviously it's not 100% _necessary_ -- there have been computers in the past that used byte sizes that weren’t a power of 2)
Robin
@b0rk @palin
Oh, and I also remember "nibble copiers" from the C64 days, but didn't know enough at the time to know why working in 4 bits would be an advantage to getting past copy protection over 8 bits. And haven't looked it up since.
Vertigo #$FF
@eythian @b0rk @palin The disk encoding uses "group coded recording", or GCR, on the disk surface. These are 5-bit codes with certain properties useful for magnetic media. Of the possible 32 binary values that have these properties, only 16 are valid. Hence, each five bit group records a single nybble.
If you copied a disk one group at a time, you could overcome some copy protection schemes that rely upon invalid group representations.
Software that focused on copying disks this way are called nybblers because they focus on copying data one group at a time, in effect performing error correction at the individual group (hence, nybble) level.
🇺🇦 haxadecimal
@b0rk Many early computers used word sizes that were multiples of six bits, and used five or six bit character codes (predating ASCII and EBCDIC). 36 bits was common for big computers, 18 for medium, and 12 bits for small.
clacke
Vertigo #$FF
@adventofcomputing @tobyjaffey @b0rk We do, though. See this document by the 8086's inventor, Steve Morse.
Intel Microprocessors: 8008 to 8086 - SteveMorse.org https://stevemorse.org/8086history/8086history.pdf
Advent of Computing
@tobyjaffey @b0rk I've been summoned! I'm also a little perplexed by the standardization on 8-bit bytes in the x86 architecture specifically. Since the 8086 was initially a stop-gap project we don't have super super good details on it's development history. That said, I think the place to look is the 8008. That chip was designed as part of a contract for CTC, which was a terminal company. The 8008 would have powered the Datapoint 2200. That machine called for 8-bit bytes:
https://history-computer.com/Library/2200_Reference_Manual.pdf
Toby Jaffey 🏳️🌈
@b0rk On the "4 bits is too small", Baudot's 5 bit code was the standard for the telegraph. 8 bits is the next smallest power of 2. https://en.wikipedia.org/wiki/Baudot_code
@adventofcomputing Had a great episode on the history of it, https://www.youtube.com/watch?v=ixPDr5mX8dg
Graham Sutherland / Polynomial
@b0rk another reason for using power of two sizes is that clock dividers that work based on halving are trivial to design (a flip flop whose input is driven by the inverse of its output will produce a square wave at half the clock frequency you drive it at), so if you think about designing a circuit where some data is sent serially (bit by bit) down a single wire and you want to trigger some event at the end of each byte, you can use three /2 clock dividers in series to get /8.
Graham Sutherland / Polynomial
@b0rk there are lots of cases where computers perform operations around power-of-2 sized "objects" (e.g. page table translation, bitmaps, etc.) that rely on a single element being representable by a single bit inside a packed array. so having the base elements also fit within that sizing scheme makes a lot of sense.
Graham Sutherland / Polynomial
@b0rk the bus size shouldn't matter too much - many things are non-power-of-2 sized in terms of data and address buses inside processors and other electronics.
one benefit of power-of-2 sizes is that the size of the element in bits itself encodes as a single set bit. 8 in binary is 0b00001000. 16 is 0b00010000. if you had a 10-bit byte the size (10 in decimal) would encode as 0b0000001010. so reasoning about the size of things in a computer for non-power-of-2 bytes gets trickier.
Graham Sutherland / Polynomial
@b0rk if you're struggling to visualise the clock divider thing I can smash together a quick Falstad sim to show it off if you want? (It's web based, I just post a link and you don't need to install anything)
Graham Sutherland / Polynomial
@b0rk but if your data is 10 bits long then you need a way to divide the clock rate by 10, which is doable but far less simple, and leads to more problems with things like propagation delay.
Dan Lyke
@b0rk hey, just a note to tell you I appreciate these cool trips down memory lane. It's fun to have all of these reminders about what I used to find so compelling about computing.
Julia Evans
@gsuberland no pressure -- i really appreciate all the answers, thank you!
Graham Sutherland / Polynomial
@b0rk I'm heading to bed now but will put something together in the morning :)
Julia Evans
@gsuberland yes I'd love that!
Willem Van den Ende - Writing
@SteveBellovin @b0rk @oclsc That is a great and loving obituary. Thank you. (I've never met Fred Brooks, but have benefited from his writings).
And a great answer. Didn't know that about 8-bit.
The Blue Wizard
@vertigo @b0rk @palin In early days of electronic computing there used to be having various sizes like 12 bit, 36 bits, etc. Then IBM rolled out the seminal computer called S/360 which standardized on the 8-bit, 16-bit, etc. as groups of bits. That has a huge influence on the subsequent development The 4004 used 4-bit since it was hard to develop more complicated circuitry on silicon back then; besides it was meant for use in calculator (speed isn't important...humans are slow! ?)
See https://en.wikipedia.org/wiki/IBM_System/360#Influential_features
Vertigo #$FF
@b0rk @palin For the same reason we talk about "digits" in decimal numbers.
A "bit" in a binary (base-2) number can be 0 or 1.
An "triad" in an octal (base-8) number can be 0, 1, 2, ..., 6, or 7.
A "digit" in a decimal (base-10) number can be 0, 1, 2, ..., 8, or 9.
A "quartet" or "nybble" in a hex (base-16) number can be 0, 1, 2, ..., E, or F. Nybble is "half a byte," making it something of a joke.
An alternative name for what we call a "byte" today is, as you'd probably guess, "octet."
Basically, these terms are ways of communicating two pieces of information at the same time: that you're talking about a single digit in some numerical base of some kind, and what the base actually is.
🇺🇦 haxadecimal
@clacke @vertigo @tobyjaffey @adventofcomputing @b0rk The PDP-10 was a reimplementation of the PDP-6, with only minor architectural changes, so it's a pre-IBM/360 architecture, not post.
clacke
Matt Austern
@b0rk I wonder to what extent it's less about some specific engineering trade-off than just because people like round numbers. Numbers like 8 and 32 sure feel rounder and more natural to me than numbers like 7 or 11 or 36.
Vertigo #$FF
@jens @b0rk @palin According to ISO standards, an octet is explicitly defined to always be eight bits.
I'd love to know the context in which "a-group-of-eight" (octet) isn't. ;)
(Brought to you by the same people who made December the 12th month.)
Jens Finkhäuser
@vertigo @b0rk @palin Ironically, it's been pointed out to me that in RFCs, an octet isn't guaranteed to be 8 bits. A good RFC should clarify that.
I forget the context. There's at least one RFC where it matters, it seems.
CMDR Yojimbosan UTC+(12|13)
@jens @vertigo @b0rk @palin Certainly a 'byte' isn't necessarily 8 bits (qv https://retrocomputing.stackexchange.com/questions/15512/did-any-computer-use-a-7-bit-byte) but I can't see "octet" being anything except 8, IIRC that's the whole reason the word was coined.
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.