25

Around the early to mid '90s it seems there was a trend for high-end workstations running some form of Unix, and running a RISC or at least some kind of non-x86 architecture. For example:

  • Sun SPARCstations
  • NeXT workstations
  • SGI workstations
  • HP workstations
  • Sony NEWS

What started them all, and why did they die out?

14
  • 5
    Isn't that a rather wide unspecified question about half of the industry over more than a while decade, regarding company policies, OSes, machines and CPUs?
    – Raffzahn
    Feb 1 at 4:39
  • 8
    While my other comment concerns more the variety of architectures that have decreased. The reason why the specific Unix/Workstations have disappeared has more to do with the performance gain of PC and that became strong enough, mass market software (AutoCAD f.ex)? Another factor was also the professionalisation of Linux which killed some of Unix niches. Feb 1 at 7:16
  • 6
    Linux also happened. In the beginning of the 90's I had a software catalog with software for PCs, and there were several kinds of Unixes listed. After '95 these started to die out.
    – chthon
    Feb 1 at 9:17
  • 9
    @OmarL Fear, Uncertainty and Doubt. In a nutshell: rumours that Intel would stop Itanium made Intel stop Itanium....
    – tofro
    Feb 1 at 9:54
  • 6
    @ThorbjørnRavnAndersen, NeXT did not become Apple. Apple predates NeXT by years, Steve Jobs having founded the latter after he was forced out of the former. Apple did later buy NeXT, and use the technology it acquired that way in the development of OS X, but that is in no way the same thing. Feb 1 at 15:53

9 Answers 9

34

It's obviously not practical to give the whole story of many companies from start to finish in a single answer. But, generally speaking in the 1980's, the computer industry was much more fractious than it is today. Many computer companies sold completely incompatible systems. There wasn't much of a single standard for operating systems or instruction sets. RISC was an emerging technology that gave excellent performance compared to 68K or x86 at the time. UNIX was sufficiently mature to be appealing to many customers, and it was easy to license it from AT&T to get started building a new OS if you were launching a computer company.

Similar market forces drove a lot of companies on a path of convergent evolution to make high end workstations that sold for more than desktop PC's, using RISC processors and a UNIX OS. As UNIX grew in popularity, it hit a critical mass in that market and became the de facto standard. It got explicitly standardized with POSIX in 1988, which meant a lot of software was easily interoperable and portable between UNIX systems. UNIX was far more capable than typical PC operating systems at the time like DOS or Classic MacOS

As time marched on, x86 got faster. It was the mass-market processor, so it got the most R&D budget compared to the high end RISC chips, even though individual RISC systems were quite expensive. Economy of scale is a powerful thing. PC operating systems got better. Mac OS X ultimately is UNIX. Linux is obviously very similar. And Windows NT isn't exactly the same thing as UNIX but there wasn't anything important that it couldn't do well enough. So there stopped being any reason to spend a bunch of money on a "high end" OS or CPU because the low end was either nearly as good, or in some cases much better. And the age of the UNIX workstation ended not with a bang but a whimper.

20
  • 12
    "fractious": I do not think that word means what you think it means. "Fragmented" is probably what you wanted.
    – JeremyP
    Feb 1 at 8:42
  • 10
    @JeremyP, From my experience of working in the computer industry starting in the early 80s, "fractious" is the perfect description.
    – Mike
    Feb 1 at 14:14
  • 2
    Indeed, in the early 1980's the low end of, say, SGI's line was much more powerful than a PC. As time went on, the PC overtook the low end workstations (price, performance, or both) which ate into the total sales (and thus product development budget). Workstations were niche machines in many ways, serving a limited market. That market just seemed large-ish compared with with the business PC market.
    – Jon Custer
    Feb 1 at 15:30
  • 2
    @ThorbjørnRavnAndersen Video games drove the price of graphics processors down, allowing them to be incorporated into less expensive PCs, so you didn't need to buy high-end workstations for fancy graphics applications. Macs became the computer of choice for this.
    – Barmar
    Feb 1 at 15:34
  • 3
    A modern tablet is comparable to a whole early 90's server farm. A super low end off-brand burner cell phone or smart watch today is wildly more sophisticated than a super high end early 90's workstation. In the early 90's a single 100 MHz CPU with 16 MB of RAM, and an 8 bit framebuffer that couldn't do hardware texture mapping and a hard drive that couldn't do 100 IOPS cost as much as a decent car.
    – wrosecrans
    Feb 1 at 16:55
24

I can give a limited personal perspective on this because I was there! It's limited because I can only talk about what happened in my company, though that was a large multinational and is probably fairly typical.

When I started at Unilever Research in 1986 we all used a Vax 11/780 mainframe. This was a pain for big calculations like finite element hydrodynamics because you had to run the jobs overnight and they always crashed due to a silly error forcing you to wait another 24 hours before you could try again.

So when in the early 90's the unix workstations started appearing our engineering guys greeted them with open arms. They weren't as fast as the Vax, but they were orders of magnitude cheaper and you could have one on your desk. It didn't take long before the engineering department had dozens of the things.

But some time around 1993-4 I remember demonstrating to a friend that I could take his massive C code and compile and run it on my Compaq 386 running one of the early versions of Windows NT. It was a lot slower than his Dec Alpha workstation, but not an order of magnitude slower and it was a lot cheaper. By the time I left in 1997 66MHz 486s were standard and NT had settled into version 3.51, which I still consider the fastest, most stable version. The unix workstations were still faster, but the 486s were fast enough for most purposes and by this time were much, much cheaper. As a result the engineering guys had stopped buying workstations except for niche cases where you needed the most power money could buy.

I lost touch after that as I moved to a job where we only used Windows NT on Intel, but in the next five years x86 CPU speeds doubled and doubled again, and at the same time Dec ran out of R&D budget and the Alpha workstations dropped out of the marketplace. It's not hard to extrapolate this trend until the unix workstations became a niche product.

3
  • I'd lay a substantial bet that that system the engineers had to use at night wasn't managed by the most skilled because I was an expert in this area at the time and REGULARLY came to help companies like Unilever actually run jobs like what you're talking about. The biggest issue was not understanding how to optimize working set size parameters correctly to use the power the machine had properly while also lowering priority "during the day" to share the hardware better. Ah, the job offers I got were astounding! Thanks for reminding me of those memories!
    – Richard T
    Feb 1 at 22:23
  • @JohnRennie — Are you sure about the 1997 date? The 66 MHz 486 was high-end in 1992, soon to be replaced by the Pentium; and NT 4 came out in 1996. Apr 12 at 11:08
  • @MichaelGraf it's a long time ago, but I'm fairly certain I was still using a 486-66 when I left Unilever in 1997, but I got a Pentium at my next job. Back in those days big companies depreciated computers over five years, though I think this changed to three years at about the time I left. Apr 12 at 15:02
9

Innovators dilemma.

The workstation vendors catered to high-end customers with expensive products, while ignoring that cheap high volume mass market products that were rapidly catching up in raw performance (due to volume scaling efficiencies in manufacturing, and eventually R&D). Thus a significant difference in price-performance appeared; and the number of customers willing to pay a large amount for a shrinking difference in performance decreased to where the workstation business became unprofitable. And unprofitable businesses become far less capable (if capable at all) of the increasing R&D costs needed to keep up with competitive market forces (indirectly related to Moore's law).

Consumer devices, such as an Apple Watch, et.al., are now capable of well over 1 Gflop of IEEE float double scientific compute (plus GPU rendering), vastly more than a Sun, SGI, or HP workstation of mid to late '90's vintage.

5
  • And said device probably runs a kernel based on Mach or Linux on a RISC CPU.
    – Davislor
    Feb 1 at 20:01
  • 2
    I don't know, my SGI O2 with 300MHz R10000 still impresses people with some of the 3d demos, especially when I tell them how old it is! And if I really want to impress them I break out my Octane 2 with dual 400MHz R12000s, I just don't want to have to pick it up and carry it very far.
    – Glen Yates
    Feb 1 at 22:17
  • 4
    I know, because I actually managed to run some of my same FP benchmark code on R10k servers (when I worked at SGI and MIPs), and on my Apple Watch series 6 (dual 64-bit). The Watch is faster.
    – hotpaw2
    Feb 1 at 22:33
  • They didn’t ignore it. SGI introduced an x86 line. Sun supported x86. Feb 3 at 20:51
  • I don't agree about that . Sun did wanted to sell a lot of decently priced WS though Sun killed of their 80s 386 serie (after the intro of the SPARC workstations) and then a number of years later did the new x86 series using intel and amd processors. Feb 15 at 13:54
8

"I was there," too.

First, to actually answer the question posed as nobody else has!

What happened to those workstations is that people like me and several friends of mine collected up the old hardware and put them to use in our smaller, sometimes personal, sometimes small business internal network environments doing jobs like serving as firewalls, file servers, etc.

My own favorite use of them was as compile and testing systems so that I could know my code worked on, say, SunOS / Solaris, and MacOS, etc, etc. I still have some of these old vintage systems. They tend to sit powered down until there's some specific need for "porting."

There were also some projects to harness the old stuff into giant computing clusters, such as The NOW Project at UC Berkeley - Network of Workstations. ...I happened to inherit part of that cluster when the project's funding ran out.

THAT'S what happened to them!

As for the market, most others have already covered that. I'll add that Windows NT is VMS (yes, the very same VAX/VMS that ran on the Vax 11/780 and follow-on systems), just ported to the x86 hardware. It got to Microsoft via Dave Coulter (sp?) who got permission to start with it in his "DEC West" venture in the Seattle area - I'm sure some others know the story better than I do, as I was just working for DEC at the time in another region. I understand that the biggest problem for the team was to figure out how to do locking correctly as the former hardware had a single instruction (machine "op code") for testing the state of a bit, set or cleared, and branching as a result or not. This made it a perfect tool for writing a lock manager. ... Unfortunately, the Microsoft people really screwed up the user interface in my view, but, well, they were trying to make it compatible with the former junk they called Windows 3.x (or whatever it was - everyone I knew called it "Blue Screen", or BSOD - blue screen of death). And yes, I acknowledge my bias; in my view DOS was already obsolete the day Gates hawked it to their very first customer as I had ALREADY written a real-time multi-tasking operating system for the Motorola 6809 chipset for TANO Corporation's "Outpost" computer.

DEC really blew it with their marketing / licensing of the Alpha chip. ... I have a dud example (a production chip that failed testing that they made into plaques for presentations) that is super fascinating...

1
6

The rise of fast 64-bit x86-compatible processors, I would argue, was the beginning of the end for the diverse proprietary processor hardware and the associated specific Unix flavours. Throughout the '90s if you asked application developers why they couldn't port their packages to x86 platforms, "memory" would likely be among the top answers, with "disk speed" a close second, along with the capabilities of Unix compared to the DOS/Windows features of the day. As the x86 platform moved to 64-bit addressing and as PC bus speeds (and thus disk and memory performance) caught up with contemporary workstations, application categories began to merge ("desktop publishing" into "word processing", for example) and the cost-benefit of Intel/AMD systems from many competing vendors just took over the world.

As to what became of them, well, they've been purchased in one form or another and mostly wound down as functioning enterprises. Sun, of course, merged into Oracle and of the legacy CPU platforms, SPARC is still very much available and has some compelling use cases. Companies that send sales teams to sell small volumes of pricy systems are just completely different enterprises than the ones that push low-markup commodities on price, and the PC clone makers were way ahead. HP as it exists today might as well be called Compaq (and have let the Test and Measurement division keep the HP name). Some of the players tried to pivot into 3rd-party peripherals or integration and maintenance services. I mean, they had to try something. Some of the big brands were valued for their patent portfolio or customer base or just for their name as a foot-in-the-door in some sectors. For a while in the early 2000's you might have been pitched an X86 cluster called Silicon Graphics or an Opteron box labeled Cray. By that time the namesake original companies were long gone.

5
  • I think the bigger issue with porting to PCs was fast high-resolution graphics and X-windows capabilities.
    – Jon Custer
    Feb 1 at 19:23
  • 9
    64-bit x86 came towards the end of the demise of RISC Unix workstations, not its beginning — it was the nail in the coffin. Opteron (the first 64-bit x86 CPU) was released in 2003, and by then PCs had largely supplanted workstations. Feb 1 at 20:01
  • I have a compaq xps 333d - that ugly machine has three simm slots - but to go to the machine's max 384 MB i need to remove both 64s and replace them with unobtanium special early version 128MB simms, while the SUN SS10 i have supports 512 MB in 64 MB sticks - which is available (though reused.) Feb 15 at 13:58
  • The 90s workstations usually except for the really cheap ones (cheap sort of) supports a LOT OF MEMORY. That SS10 was i believe in the really low price ones before Solaris 2 sold with 32 MB in low spec config and 128 MB at a high spec (and faster CPU too) while the machine do support though badly (temp issues and bus speed) 4 processor machines and at a max 512 MB mem. Feb 15 at 14:02
  • The memory span was done using a lot of not necessarily populated memory slots (=costs) while the PC manufacturers saved costs only having 3 or 4 slots in the mid 90s Feb 15 at 14:05
1

Back then I was doing engineering design work using specialized design software.

The hardware options were:

The Default - an MS DOS computer with WISIWIG. The graphics weren't brilliant, but usable and monitors were usually 13 inches in size, CPU speeds for high volume number crunching was passable, MS DOS computers could use high capacity hard drives, but they had to be partitioned into 33 MB chunks.

Unix Workstations - had better graphics and larger screens, up to 23 inches, CPU speeds were better and hard drives only needed one partition.

The downfall of Unix workstations were the high cost and the availability of software that could be used on them.

Additionally, companies started asking "why does one person need two computers to do engineering design, why can't one be used?".

A typical engineering design project would involve using the Unix workstation to do the number crunching and graphical design, but the MS DOS box would be used for financial and other calculations using spreadsheets (Lotus 123, Quattro Pro, Excel) and to write the report (Word Perfect, Word). Any databases used where also on the MS DOS box (dBaseII or Access). We didn't have spreadsheet or word processor packages for Unix workstations.

Any data transferred between the two computers was done using floppy disk. However, because of the issue of carriage return and line feeds at the end of lines in text files between Unix and DOS, prior to transfer the format of the file had to be changed using a Unix to Dos conversion, or vice versa.

When MS DOS boxes were able to have higher technical specifications and when specialized engineering design software could be ported to MS DOS boxes, Unix workstations were abandoned. They were too high cost and so some extent under utilized.

1

Heh... You may want to watch "Revolution OS" which details the how's and whys. Back when I was in college in the last '80s, we usually had terminals into a Minicomputer like a DEC VAX7800/8800. Beyond that, people had 8-bit computers or dumb terminals. Having one's own SPARCstation 5 or 10 was still rare, people would use PC/XTs to talk to an AT&T System V or BSD derived system like Ultrix/Xenix and similar.

The 16-bit era was truly short. Like from 1985 to 1990 as Packard Bell was moving a lot of 386/486 stock along with other 3rd party makers like Gateway, Compaq, and Dell. There were Macs and PC-DOS/Win3.11 setups prior to 1995 with TCP/IP stacks. On the PC side, many were using Ham radio software from KA9Q as ham radio operators had their own network space from the very beginning of the Internet itself.

Frankly, the Unix workstation market was killed by Open Source: Linux and Free/Open/NetBSD. At a price tag of $8000-$35000 for a workstation (like an SGI Indy R10k) compared to commodity/surplus PC/AT with 386 or 486 they were suitable enough to run as a workstation. Think how prevalent the "Computer Shopper" was during that time with Mail-Order PCs during the height of the .com era.

The Pentium and Pentium Pro of the late 1990s further lead to better performance and graphic capabilities with ATI Rage128 and S3 Virge graphics. Most Unix workstations couldn't do serious 3D graphics.

Even DEC Alpha DS10s were PCI-based but the lack of extensive graphics didn't make it a powerhouse unless it had graphics processing boards that cost 5-digit sums of money to pull off arcade levels of frame rate or rendered to a FrameStore. Computer animation even today relies on data centers called "Render Farms" to frame by frame render and properly light each frame of animation being generated for movies like Pixar's Toy Story.

NeXT and eventually, Sun Microsystems did see the light early in the mid-nineties and took steps to port their notable operating systems from Motorola 68000 or MIPS-based platforms to Intel x86. As did versions of NT4.5 and Windows 2000 for Alpha and HP PA-RISC processors.

At the end of the day, just like how Minicomputers of the early 90s made by Unisys, DEC/HP, and others turned into commodity hardware that made up the last 20 years that are now turning into specialized data center hardware bringing the network edge right to the top of the server rack and the rest of the volume essentially standard hardware or customized versions of the same by OEMs.

The evolutions in datacenters have come from better power efficiency and the ability to increase computational capacity per U of rack space. This requires having cheap power and backing infrastructure to force cool air into the grasp of higher volume server fan inlets which force air through rack equipment using extremely high RPM internal fans and is vented enmass back out to chilling spaces.

Whether or not it was 32 or 64 bits, that didn't matter as much as memory and ease of coding for the architecture were bigger blockers. Much like doing parallel computing to take advantage of Cray Supercomputers which were 386/486 processors being configured in wide parallel over 128 processing dies. Most software is still single-threaded today with software now being rewritten and updated to take advantage of multi-threads now in current processors better.

But all the workstation makers of old became companies like Hard Drives NorthWest and Dell. What made them valuable was their patents and IP which were resold to the big companies like HP, Oracle, Apple, and others.

That's the way it goes.

0

There is also the maintenance aspect. When you have 1000+ employees and each of them have an x86 desktop, they can be looked after by a few souls for a small relative cost. If you have a few employees running highend workstations, you may still need a few souls to look after them, however the relative cost then becomes high. Therefore customers would be looking to move them to cheaper x86 systems. The cost of an employee does tend to rather outway the cost of even a highend workstation.

11
  • 7
    With automated configuration tools and scripts, a few hardy souls can look after a surprising large number of UNIX systems. Anecdotally, it takes fewer people to manage a UNIX/Linux systems than it does Windows systems.
    – doneal24
    Feb 1 at 22:00
  • @doneal24 Lets say 1000 systems looked after by 10 folk, compared to 10 systems looked after by 2. 1:100...1:5. There is no comparison. And remember, this is back in the 90's. I think most companies would be looking to lower their wage bill by getting rid of the workstations and those 2 employees.
    – Bib
    Feb 1 at 22:04
  • 3
    IME companies and institutions (e.g. universities) with large fleets of workstations didn’t need many people to look after them; in some cases, fewer than the similar fleets of PCs in the same companies or institutions... Feb 1 at 22:19
  • @bib I am talking about 2-3 people managing 200-500 UNIX/Linux systems. I don't know where you got the idea that my "surprising large" number was around 10. Recalculate the ratios with 1000 Windows systems with 10 administrators compared to 500 UNIX systems managed by 2 people.
    – doneal24
    Feb 1 at 22:42
  • How do 1000 employees sitting at mass-market, i386 and i486 PCs running 16-bit Windows do the work of even one engineer sitting at a high-end, Unix-based, CAD workstation or, of even one film editor or FX artist sitting at an SGI Indigo worstation? Feb 2 at 1:58
0

Microsoft Office

The thing that gradually doomed the UNIX workstations was Microsoft Office. Corporate managers really wanted their engineers to be able to read and write Word, Excel and PowerPoint documents, and were annoyed that engineers couldn't do that on their workstations, and even worse, weren't concerned about it. Giving them a second computer for Office was obviously expensive, so when Windows PCs became capable of running high-end engineering software, that was what all the customers demanded.

This need was understood long before Windows PCs actually took over.

Pentium Pro

Before Intel shipped the Pentium Pro in November 1995, all the RISC Unixes had major performance advantages over PCs. It was reasonably straightforward for an ISV to offer the same software on several of them, since they all used Unix, similar dialects of C and C++, and X-Windows for GUI.

The Pentium Pro was the first mass-market processor to use the modern technique of breaking down more complex instructions into micro-operations and having a "pool" of them that Functional Units could grab operations out of to execute when the data they needed became available. The MIPS and SPARC processors of the time suddenly looked vulnerable. One company that had been set up to offer fast MIPS-based Windows NT machines abruptly changed to using Intel.

The original Pentium Pro's impact in the general Windows market was limited, because it ran the 16-bit code that was common in Windows 95/98 rather slowly. That was utterly irrelevant for competing with UNIX workstations, where all the relevant software was fully 32-bit.

The advent of 64-bit

This actually started before the Pentium Pro, but didn't become important until the price of memory was low enough that having multiple GB of RAM was sensible. It marked an important change: memory limits became a soft cap rather than a hard cap.

With 32-bit addressing, using more than 4GB of address space in a process (with inevitable subtractions for the OS) was annoyingly hard work. People who'd had to work with EMS and XMS memory under MS-DOS were reluctant to use PAE or bank-switching: it was clear that 64-bit addressing would be a better solution, and that it was on the way.

So with 32-bit addressing, there was a hard limit on your "document" size. This wasn't usually a traditional document, but a CAD, CFD, or other kind of "model." With 64-bit, you could make bigger and bigger models. They got slower when you exhausted the physical memory of your machine, but you could buy more RAM.

The Itanium fiasco

The long-term significance of Itanium was that it disrupted the development of the main RISC architectures:

  • MIPS development stopped for several years, and never caught up again.
  • Alpha was killed off.
  • PA-RISC was replaced.
  • SPARC and POWER only had brief slowdowns, which is part of why they're still going.

AMD did a good job

One of the reasons Itanium was so unlike x86 was so that it could be patented up to the eyeballs, and none of it would be subject to AMD's x86 license. AMD could not build Itanium-compatibles, so they had to design something to compete with Itanium. They did that well, and Intel made several mistakes with Itanium design and product management. Intel ended up having to build AMD-compatibles.

Microsoft enforced cross-compatibility

Apparently - this is from a senior AMD person in 2005 - Intel wanted to build a 64-bit x86 that was not AMD-compatible. Their intention was to force software vendors to do separate builds for Intel and AMD, in the hope that they would not bother with AMD software, and drive AMD out of the market.

Dave Cutler of Microsoft responded that they already had 64-bit Windows working nicely on AMD64, and while they could not prevent Intel making their processors incompatible, they could and would refuse to offer Windows for such processors. Intel had to back down.

3
  • From what I understand, a major problem with Itanium is that while a multi-core processor can be made much faster if it can make certain assumptions about memory-access patterns in multi-threaded programs, making programs that aren't designed around such assumptions work correctly on such architectures will often require adding many expensive memory barriers, yielding performance which is substantially inferior to what was achieved on platforms that didn't make such assumptions.
    – supercat
    Feb 24 at 18:55
  • 1
    @supercat: That's true, but the problems with memory access on Itanium are much broader than that. They apply even in a single-threaded program running by itself on a machine. I suppose I could write a question and answer about that if it was interesting. Feb 24 at 19:01
  • I don't know to what extent people would find the details interesting, but I think it's important to note that Itanium and Alpha could run faster than competing architectures when running code that was designed from the ground up around that architecture's strengths and weaknesses, but could not perform well with a lot of code that was designed for architectures which had a stronger coupling between cores.
    – supercat
    Feb 24 at 19:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.