The curious case of the Unix workstation layout(thejpster.org.uk)
thejpster.org.uk
The curious case of the Unix workstation layout
https://thejpster.org.uk/blog/blog-2025-07-19/
42 comments
I suspect a big factor was PSU design. Even large OEMs contracted it out, and there probably reached a point where manufacturers started to offer catalog items in the 200-400 watt range where the primary size constraint was the 80mm exhaust fan. Once that's the constraint, it's going to look a lot like LPX, and you get a reinforcement factor around specific dimensions when vendors say "If we use the exact same screw positions we can compete as a drop-in replacement"
It's interesting that PCs adopted the LPX PSU but not the cases and motherboards. I had always seen LPX described as a bit more proprietary than (baby-)AT designs-- you could fit anyone's mainboard in a generic AT case, but even a HP "LPX" mainboard and riser card might not fit in a Packard Bell case.
It's interesting that PCs adopted the LPX PSU but not the cases and motherboards. I had always seen LPX described as a bit more proprietary than (baby-)AT designs-- you could fit anyone's mainboard in a generic AT case, but even a HP "LPX" mainboard and riser card might not fit in a Packard Bell case.
These days the majority of computers are designed and manufactured by a small number of OEMs too.
> Early Alphastations were VME based machines.
No, they most emphatically were not. The early 3000 series, like the MIPS based DECstation and VAX based VAXstation, were based on Turbochannel[1], which was DEC's primary bus technology at the time. Later Alphas used PCI as their primary bus. There were Tubochannel to VME adapters, and PCI to VME adapters, but VME was never a primary bus used in alpha.
1: https://en.wikipedia.org/wiki/TURBOchannel
No, they most emphatically were not. The early 3000 series, like the MIPS based DECstation and VAX based VAXstation, were based on Turbochannel[1], which was DEC's primary bus technology at the time. Later Alphas used PCI as their primary bus. There were Tubochannel to VME adapters, and PCI to VME adapters, but VME was never a primary bus used in alpha.
1: https://en.wikipedia.org/wiki/TURBOchannel
Im just going to make sure at least one post here mentions S-100 bus
https://en.wikipedia.org/wiki/Altair_8800
https://en.wikipedia.org/wiki/Vector_Graphic
https://en.wikipedia.org/wiki/North_Star_Horizon
https://en.wikipedia.org/wiki/IMSAI_8080
Thanks for coming to my ted talk
https://en.wikipedia.org/wiki/Altair_8800
https://en.wikipedia.org/wiki/Vector_Graphic
https://en.wikipedia.org/wiki/North_Star_Horizon
https://en.wikipedia.org/wiki/IMSAI_8080
Thanks for coming to my ted talk
I made a living porting Unix to new hardware in the mid to late 80s so all the busses of the day came across my desk, and designed graphics card silicon in the 90s
S-100 was very much history by the time the boxes in the article were designed (5-10 years before) VME and Multibus were the first generation workstation busses, PCs had ISA->EISA, Macs had NuBus/NuBus90 - all of them converged on PCI once chips with enough pins were packaged cheaply enough (plastic rather than ceramic - 200+ pins)
S-100 was very much history by the time the boxes in the article were designed (5-10 years before) VME and Multibus were the first generation workstation busses, PCs had ISA->EISA, Macs had NuBus/NuBus90 - all of them converged on PCI once chips with enough pins were packaged cheaply enough (plastic rather than ceramic - 200+ pins)
SGI Indigos (not Indigo^2) were SGI's own GIO32 bus, not VME.
Early Alphastations were assuredly not VME! They were DEC's own Turbochannel.
Early Alphastations were assuredly not VME! They were DEC's own Turbochannel.
Indigo R3000 was VME-based, a derivative of the 4D/35 and firends
Indigo R4xxx was a completely different architecture shared with the later SGI Indy
Both had GIO32 slots
Same case, same slots, totally different architectures inside
Indigo R4xxx was a completely different architecture shared with the later SGI Indy
Both had GIO32 slots
Same case, same slots, totally different architectures inside
...and Turbochannel was originally the bus for the MIPS R3000-based Ultrix workstations; the Alpha firmware needed to be able to interpret MIPS machine code to speak to TC expansion cards.
Somewhere, I have a Turbochannel FDDI card taken from an AlphaStation 3000-an obsolete I/O adapter for a forgotten network protocol, running on a defunct CPU, with firmware written in the native language of an even more defunct CPU.
Somewhere, I have a Turbochannel FDDI card taken from an AlphaStation 3000-an obsolete I/O adapter for a forgotten network protocol, running on a defunct CPU, with firmware written in the native language of an even more defunct CPU.
You can get a TURBOchannel adapter for the VAXstation 4000/60 or /90 and it also has a MIPS emulator for running TURBOchannel card ROMs.
It makes sense, if one considers them the evolution of the Apple II and the IBM PC.
The IBM had the motherboard on the left, along with the expansion cards, and the drives and power supply on the right. The AT continued this.
Clones wanted mostly compatible cases, and motherboards wanted to be mostly compatible with IBM cases and clone cases.
Then we had Amigas like the Amiga 3000, which had a similar layout but a riser to take horizontal expansion cards.
While some more bespoke PCs had vertical risers, most PC cases in the early to mid '90s were large. It was the machines we paid a bit more for that made being smaller in to something a bit premium.
While taking apart my Amiga 3000 is a bit of work, the design is absolutely wonderful, and more than once I thought about the design of it compared with later machines like the Sun Ultra 5, the Motorola StarMax (PowerPC Mac clone) and others.
The IBM had the motherboard on the left, along with the expansion cards, and the drives and power supply on the right. The AT continued this.
Clones wanted mostly compatible cases, and motherboards wanted to be mostly compatible with IBM cases and clone cases.
Then we had Amigas like the Amiga 3000, which had a similar layout but a riser to take horizontal expansion cards.
While some more bespoke PCs had vertical risers, most PC cases in the early to mid '90s were large. It was the machines we paid a bit more for that made being smaller in to something a bit premium.
While taking apart my Amiga 3000 is a bit of work, the design is absolutely wonderful, and more than once I thought about the design of it compared with later machines like the Sun Ultra 5, the Motorola StarMax (PowerPC Mac clone) and others.
> like the Sun Ultra 5
The Ultra 5 (desktop) and Ultra 10 (tower) were a cost cutting exercise that put an UltraSPARC IIi (2i) onto what I think was an ATX form factor motherboard. It used ATA drives, USB keyboard and mouse, a VGA port, etc. This was an act of desperation from Sun, not an example of their best engineering.
That said, compared the performance of a $3500 Ultra 10 with 512 MiB of RAM to $10k+ Sun Ultra 30’s and HP C180’s, each with 128 MiB of RAM. These prices were after applying significant edu discounts. The heftier sheet metal, SCSI drives, and nostalgia did not allow these traditional UNIX workstations to touch the performance of the much cheaper Ultra 10 with 4x the RAM.
The Ultra 5 (desktop) and Ultra 10 (tower) were a cost cutting exercise that put an UltraSPARC IIi (2i) onto what I think was an ATX form factor motherboard. It used ATA drives, USB keyboard and mouse, a VGA port, etc. This was an act of desperation from Sun, not an example of their best engineering.
That said, compared the performance of a $3500 Ultra 10 with 512 MiB of RAM to $10k+ Sun Ultra 30’s and HP C180’s, each with 128 MiB of RAM. These prices were after applying significant edu discounts. The heftier sheet metal, SCSI drives, and nostalgia did not allow these traditional UNIX workstations to touch the performance of the much cheaper Ultra 10 with 4x the RAM.
Hmm...makes me wonder if a PA-RISC HP 9000 712 is a better NeXTStation than a NeXTStation is today, in terms of longevity, supportability, performance, etc.
I guess its missing the DSP and fancy printer interface of the NS, maybe the overall sound quality.
I guess its missing the DSP and fancy printer interface of the NS, maybe the overall sound quality.
Unlikely to be better in supportability terms, since we are talking about a 30 year old OS on 30 year old hardware that wasnt the primary development target.
However, in terms of build quality, I would take the old HP 9000s over NeXT cube any day. Those old HPs were engineering workstations built by an HP that at the time still cared about engineering.
Jobs and the NeXT crew built beautiful things, but at that time they didnt have the decades of experience that HP did. (I'd say the same about old RS/6000s by the way.)
A trivial example is how when you open up the back, engineering workstations did not slice up your hands on stamped metal frames and risers. Whereas PCs of the time generally did. (Compaq was a notable exception, but you paid a price premium for it.)
However, in terms of build quality, I would take the old HP 9000s over NeXT cube any day. Those old HPs were engineering workstations built by an HP that at the time still cared about engineering.
Jobs and the NeXT crew built beautiful things, but at that time they didnt have the decades of experience that HP did. (I'd say the same about old RS/6000s by the way.)
A trivial example is how when you open up the back, engineering workstations did not slice up your hands on stamped metal frames and risers. Whereas PCs of the time generally did. (Compaq was a notable exception, but you paid a price premium for it.)
> Those old HPs were engineering workstations built by an HP that at the time still cared about engineering.
I agree. They are heavy and very strong, and chassis is very stiff. That mattered given where some of these ended up being used and how they were treated, despite the cost.
I agree. They are heavy and very strong, and chassis is very stiff. That mattered given where some of these ended up being used and how they were treated, despite the cost.
Something to keep in mind is available software for NeXT computers versus for platforms that NeXTstep got ported to (x86, PA-RISC, SPARC). Many later software written for NeXTstep and its successor OPENSTEP had “fat binaries,” meaning that they contained machine code for multiple platforms. However, there are noticeable software packages released before NeXT’s transition away from their own hardware that only have a 68k binary. I believe WordPerfect, Lotus Improv, and Adobe Illustrator fall in this latter category.
I own a NeXT Cube and three NeXTstations.
I own a NeXT Cube and three NeXTstations.
When I was an intern at Apple in late 1997, my main machine was a PA RISC workstation running OPENSTEP for Mach (can’t recall which version). It was a lot faster than the PowerPC I had on my desk so I did much of my development on that before testing on Rhapsody.
I run it on a SAIC Galaxy 1100 which is a Gecko in a MIL-SPEC portable workstation case, and it is a delight to use. It's snappy and slick. But I have a soft spot for PA-RISC anyway since working on a K250 midrange was my first job out of college.
The 712 was already a better NeXTstation than the actual NeXTstation at the time
It cost half as much and had considerably more CPU power
There is a reason NeXT decided to become a software company...
It cost half as much and had considerably more CPU power
There is a reason NeXT decided to become a software company...
Old workstations that were purposely designed to run proprietary UNIX(r) were also, coincidentally, super-proprietary.
If you could install IRIX on junk commodity hardware no one would have a reason to pay SGI $100k for one of theirs.
If you could install IRIX on junk commodity hardware no one would have a reason to pay SGI $100k for one of theirs.
This makes it sound like being able to run IRIX was what got people to buy SGI. But I think the reason why people bought $100k SGI machines was for the graphics hardware, not for IRIX.
IRIX was so far ahead of its time that the software was a real selling point. We had a bunch of different unix workstations in my PhD group, but as soon as someone got an Indy, it was SGI all the way. No-one wanted anything else.
Six of one, half a dozen of the other: The software that took advantage of the graphics hardware ran on IRIX, and your system administrators and software developers would also probably need IRIX systems, but without as much need for the graphics functionality. This was one of the reasons for the Solaris x86 port: It might still be useful for people who didn’t need a Sun to run Solaris.
Back in 1993, a $5k SGI had ~5x as much CPU power as a $10k x86 PC. Not a joke. You could get a 100 MHz R4000PC that would happily drive a display at 1280x1024 for the price of a 33 MHz 486 that would run Windows 3.x at 640x480
They worked very hard to make IRIX a competitive advantage over other UNIX vendors, but it was never a reason to buy SGI over "commodity"
If anything, IRIX was a hindrance to adoption, because UNIX was notoriously RAM-hungry, and the early 1990s had a horrible, horrible price crunch on RAM.
They worked very hard to make IRIX a competitive advantage over other UNIX vendors, but it was never a reason to buy SGI over "commodity"
If anything, IRIX was a hindrance to adoption, because UNIX was notoriously RAM-hungry, and the early 1990s had a horrible, horrible price crunch on RAM.
My first computer was a Packard Bell 486-33 with 4MB RAM, purchased in 1993. I can still remember my widowed schoolteacher mother's pursed lips as she paid $1600 - an outrageous sum of money - for it at Walmart. I don't know who was paying $10K for PCs in 1993, but it wasn't anybody I knew.
Walmart vs. IBM, Compaq, Hewlett-Packard, Apricot and the likes.
I remember this, because around that time I got my first PC(not my first computer), which was a 25Mhz 486. For 5000.- Deutsche Mark. At that time such systems from IBM, Compaq and so on sold for 15.000.- Deutsche Mark.
My system lacked the cache RAMs for the L2, and the tag RAM for these, though I could, and did plug these in afterwards, for actually not that much money(about 80 Deutschmarks).
It also had a slow and small PATA disk, on a mediocre mainboard. Just 80MB. The VGA screen I initially got with it was also mediocre at best, as was the ISA VGA in a VESA Local Bus system...
For the 15.000 of whicheveruniversalcreditunits at the time, vs. just 5000 of them, you'd gotten a more balanced system from IBM, Compaq, etc., with larger and faster disk(s), more RAM, working L2-cache, maybe even SCSI, better graphics, probably better screen, keyboard & mouse.
I remember this, because around that time I got my first PC(not my first computer), which was a 25Mhz 486. For 5000.- Deutsche Mark. At that time such systems from IBM, Compaq and so on sold for 15.000.- Deutsche Mark.
My system lacked the cache RAMs for the L2, and the tag RAM for these, though I could, and did plug these in afterwards, for actually not that much money(about 80 Deutschmarks).
It also had a slow and small PATA disk, on a mediocre mainboard. Just 80MB. The VGA screen I initially got with it was also mediocre at best, as was the ISA VGA in a VESA Local Bus system...
For the 15.000 of whicheveruniversalcreditunits at the time, vs. just 5000 of them, you'd gotten a more balanced system from IBM, Compaq, etc., with larger and faster disk(s), more RAM, working L2-cache, maybe even SCSI, better graphics, probably better screen, keyboard & mouse.
But they were really, really good. An Indigo² Solid Impact R4400SC 250 was my main PC from 1998 until I left the US in late 2003. Built like a tank. Incredibly smooth in operation. Rang rings around my iMac G4, which ended up just sitting on top of the SGI's Trinitron running iTunes. I loved that machine, and I will go retrieve it in a few months.
Now the Fuel, of the final MIPS generation, was hot garbage. Early-2000s PC crap plus SGI does not equal awesome. Fast, but flaky, stuttery, and with too many graphical artifacts for the price. Maybe the V12 was better than the V10. I don't care to find out.
Now the Fuel, of the final MIPS generation, was hot garbage. Early-2000s PC crap plus SGI does not equal awesome. Fast, but flaky, stuttery, and with too many graphical artifacts for the price. Maybe the V12 was better than the V10. I don't care to find out.
I've got a 900MHz Fuel with a V12 DCD. It runs fine, though I never ran a V10 in it. Biggest problem with it is the power supply so I find myself using the Indy more these days.
I remember it roughly like this:
Before the PC era computers were mostly integrated in the keyboard (many 8-bit home computers), the monitor (Apple Lisa) or both (Commodore PET).
Then they became more or less flat boxes on the desktop. Early PC's were like that (the bulky version) but also early workstations like Sun's SPARCstation (the more elegant version). They were meant to put a monitor atop.
Over time the boxes got bigger and louder and the monitors got bigger which made this design impractical. Some people flipped the boxes, put them under their table and the tower was born. Not long and professionally made tower cases appeared.
Over time the bulky towers got smaller and we had Midi- and Mini-Towers on the PC side and things like the Sun Ultra 24 or the SGI O2 on the workstation side. These could be put on the table again but this time next to the monitor and not below it.
Before the PC era computers were mostly integrated in the keyboard (many 8-bit home computers), the monitor (Apple Lisa) or both (Commodore PET).
Then they became more or less flat boxes on the desktop. Early PC's were like that (the bulky version) but also early workstations like Sun's SPARCstation (the more elegant version). They were meant to put a monitor atop.
Over time the boxes got bigger and louder and the monitors got bigger which made this design impractical. Some people flipped the boxes, put them under their table and the tower was born. Not long and professionally made tower cases appeared.
Over time the bulky towers got smaller and we had Midi- and Mini-Towers on the PC side and things like the Sun Ultra 24 or the SGI O2 on the workstation side. These could be put on the table again but this time next to the monitor and not below it.
Yes, it's interesting that SFF (small form factor) and ITX-based systems are now back to the "flat box on the desktop" style.
It’s more “shoved behind your big flat panel monitor” style, given how infrequently people use removable media or otherwise need to physically touch the machine now.
Indigo2 is just convergent evolution.
The Indigo2 family had EISA slots, so it was obliged to hold a full-length ISA card.
There are only so many ways to skin that cat: You want a small, thin desktop with quiet cooling that will also fit a full length ISA card. LPX and Indigo2 look similar because they were solving the same problem.
See also: Octane and Origin 200 vis a vis ATX and WATX. When SGI adopted 64 bit PCI, their desktop-sized systems started to look a bit like ATX, despite a radically different underlying architecture
-----------
The equivalent Sun and HP9000 kit could look quite different because they did not even try to offer EISA as an option
Later Sun and HP9000 kit, with PCI64 slots, started to converge on an ATX "look" for similar reasons to the I2 and LPX
The Indigo2 family had EISA slots, so it was obliged to hold a full-length ISA card.
There are only so many ways to skin that cat: You want a small, thin desktop with quiet cooling that will also fit a full length ISA card. LPX and Indigo2 look similar because they were solving the same problem.
See also: Octane and Origin 200 vis a vis ATX and WATX. When SGI adopted 64 bit PCI, their desktop-sized systems started to look a bit like ATX, despite a radically different underlying architecture
-----------
The equivalent Sun and HP9000 kit could look quite different because they did not even try to offer EISA as an option
Later Sun and HP9000 kit, with PCI64 slots, started to converge on an ATX "look" for similar reasons to the I2 and LPX
> Or is the LPX desktop just the crab of personal computer design, and all lineages will converge here eventually. I have no idea.
It's simply a good design for the time. Motherboards needed a bunch of stuff, none of it very tall, expansion busses were busses, so you can put one slot in the middle for a daughter board with slots fot expansion cards. Drives go on the right hand side by tradition, which makes expansion ports on the left.
Depending on how many slots you want, you can make the case taller or shorter.
It's simply a good design for the time. Motherboards needed a bunch of stuff, none of it very tall, expansion busses were busses, so you can put one slot in the middle for a daughter board with slots fot expansion cards. Drives go on the right hand side by tradition, which makes expansion ports on the left.
Depending on how many slots you want, you can make the case taller or shorter.
Let’s not forget Amiga 3000UX - a workstation released with Amiga Unix, a full port of AT&T Unix System V Release 4 (SVR4). Notable users include Free Software Foundation staff programmers who used it at MIT to help further some early development of the GNU operating system.
https://en.m.wikipedia.org/wiki/Amiga_3000UX
https://en.m.wikipedia.org/wiki/Amiga_3000UX
The 3000UX was dead on arrival
It cost as much as an early 1990s UNIX workstation but it featured the technology of the 1980s, so it was extremely slow by the standards of the day.
For the price of a 3000UX, you could buy an SGI with 10x the CPU power, or a Sun with 10x as many pixels on the display. It was a really, really bad deal. As per usual for Commodore, too little, too late.
It cost as much as an early 1990s UNIX workstation but it featured the technology of the 1980s, so it was extremely slow by the standards of the day.
For the price of a 3000UX, you could buy an SGI with 10x the CPU power, or a Sun with 10x as many pixels on the display. It was a really, really bad deal. As per usual for Commodore, too little, too late.
The early 90s were the era of (1) Pentium (2) PCI (3) multiplatform Windows NT. Most vendors switched to desktop/deskside designs based on PC bus and PC components, it would be more cost competitive (volume economics) and could run NT/Alpha, NT/MIPS, NT/PowerPC if the market went that way.
In the end it was GNU/Linux under Intel the one which ate propietary Unix workstations. By late 90's (and even more in early 00's) MESA was getting good
enough to surpass Irix machines.
Maybe at the bottom end. I was using SGI Onyx 2's in the post-production world all the way up to 2005. You couldn't run Inferno on a linux box running Mesa.
After 1999-2000 or so you could as DRI (hardware accelereted GL) was becoming a reality.
https://dri.freedesktop.org/wiki/DriHistory/
https://dri.freedesktop.org/wiki/DriHistory/
I’m not saying that SGI weren’t involved with DRI, but Inferno cost about £1,000,000 a seat, and the suite was booked out at £1000 per hour. There really isn’t any comparison between a PC running DRI graphics and an Onyx 2 with 128 CPUs and 256 GB (yes, GB, even then) of RAM, running (programmable GPU) InfiniteReality2 graphics.
It’s a mountain compared to a molehill.
It’s a mountain compared to a molehill.
Windows NT was unlikely to be in the minds of many of the OEMs of the early 90s. It was marketed against UNIX, and the UNIX market was losing more and more users to the ever-more-powerful PCs running DOS and (usually) Windows 3, with a Novell setup if you wanted to share files and printers.
It really wasn't until NT 4 that Microsoft started pushing NT Workstation for general office use.
It really wasn't until NT 4 that Microsoft started pushing NT Workstation for general office use.
Depends how you define "early 90s" but Pentium wasn't released until March 1993 and the second one in October 1994 made more of an impression.
I suspect when a company spec'd out a new design and then got all the tooling done, unless they explicit language in their contract about selling stuff made with the same tooling to others, the factories could pitch "we will do the basic case with no NRE[1]" and that was a bargain. As a result a lot of things ended up being "magically" similar in those days.
[1] NRE = Non-Recoverable Engineering which is the cost label for the engineering work to build the jigs and tooling that the factory will use to make the parts you want. Example an injected molded switch cover might cost $10,000 in NRE to make the molds that can produce 10 switch covers each and be used up to a 10,000 times. Then if you make 10,000 switch covers, you have used the mold one 1000 times and used up 10% of its lifetime. Cost of the plastic plus $1,000 (the 1/10th of the cost of the mold) are the real cost of those switch covers.