Blog Post

The search for the lost Cray supercomputer OS

In 1976, famed computer architect Seymour Cray released one of the most successful supercomputers ever made: the Cray-1, a stylish 5.5-ton C-shaped tower that was quickly embraced by laboratories all over the world. While it soon gave way to newer, faster Cray models that then faded away entirely in the ’90s due to huge cost and performance advances in supercomputing, its iconic shape and early success left a lasting legacy in the industry.

Seymour Cray and a Cray-1 supercomputer. Photo courtesy of Cray.
Seymour Cray and a Cray-1 supercomputer. Photo courtesy of Cray.

That legacy led hobbyists Chris Fenton and Andras Tantos to pose what they thought would be a simple question to answer: How can I build a Cray-1 for my desk?

A tiny replica takes shape

In 2010, Fenton, a New York City-based electrical engineer who actually works on modern supercomputers, decided to replicate the physical form of the computer. Its hardware was well-documented online, so building it came together quickly. Fenton used a CNC machine and glue to build the tower and bench out of wood. Then he painted the tower and covered the bench in pleather. The complete model is 1/10 the size of the original Cray.

Chris Fenton scale model of Cray-1
Chris Fenton scale model of Cray-1

Considering that an iPad packs far more computing power than a Cray-1, it wasn’t difficult to find a board option that could handle emulating the original Cray computational architecture. Fenton settled on the $225 Spartan 3E-1600, which is tiny enough to fit in a drawer built into the bench. Considering the first Crays cost between $5 and 8 million, that’s a pretty impressive bargain.

A lead emerges from a Minnesota basement

But while Fenton was able to replicate the architecture, he hit a wall when he began searching for software to make the Cray model fully operational. He determined none of the code from the original OS was available via the internet, so he went analog. He asked the Computer History Museum and government whether they had a copy laying around. Nope.

His first lead came via a friend who introduced him to Donald Lee, a former Cray software engineer who had “this giant 10-pound disk pack” — an early, removable medium for data storage — in the basement of his Minnesota home. Lee told me last week that he isn’t exactly sure how the disk ended up in his possession, but he might have picked it up at a garage sale or when someone was cleaning house before leaving Cray.

Cray-1 disk pack

To run the disk, Fenton needed a disk drive. He borrowed a 1970s-era one from the Museum of Information Technology at Arlington in Texas. Poor archival conditions had left the drive in rough shape, so he cobbled together a system involving a robot made from the innards of a MakerBot Thing-O-Matic. The robot slowly passed over the disk, allowing a converter to change the data from analog to digital and feed it onto Fenton’s computer.

The disk drive. Photo courtesy of Chris Fenton
The disk drive. Photo courtesy of Chris Fenton

Disappointingly, the disk pack only held factory-testing software.

A second disk pack surfaces

The publicity Fenton garnered via his documentation of the project caught the attention of Andy Gelme, an Australian software developer who once worked for Cray. He too had a disk pack.

Concerned that shipping would damage or demagnetize the disk, a friend of Gelme’s couriered it to Fenton during a planned trip to New York City a few weeks later. While the disk didn’t contain the original Cray-1 OS, it did carry the last ever version of the Cray OS, which was made for a Cray-1 successor: the Cray X-MP.

Fenton used the recovery system he developed with the disk drive to pull information off the disk. He was not able to convert it into working software, but he had begun to correspond with Tantos, a Microsoft electrical engineer, who had independently been pursuing the Cray OS.

Tantos took over working with the disk. He rewrote the recovery tools, plus a simulator for the software and supporting equipment like printers, monitors, keyboards and more. For the greater part of the last year, he arduously reverse engineered the OS from the image. Despite a few remaining bugs, the Cray OS now works.

Fenton is now in the process of upgrading his desktop machine to be compatible with the Cray X-MP OS. The two are also on the lookout for a compiler: a computer program that would allow them to write their own applications and feed them into the Cray.

“For these machines (Cray-1 or X-MP) you couldn’t really go into a store and buy an application, like you do for a PC these days. Now, you just ‘install’ Word and it runs. For these machines, everything came in source-code format and you needed to compile it before you could run it. You use the … compiler to turn it into machine code the machine could understand,” Tantos said. “That was the main way you interacted with these machines. Without the compiler, you can’t feed it that.”

If you have any information that could help Fenton and Tantos, you can contact them here.

Preserving what’s available

Both hobbyists say they tackled the project for fun, but found it to be an important exercise in the preservation of computing history. Despite the fame of these early computers, little effort was made to ensure full documentation survived.

“In some ways it’s sad, but in other ways it’s fascinating,” Tantos said. “Seeing how extremely hard it is to come by software for these early computers, it’s even more important that we preserve what is available.

When they finally are able to pair the OS with Fenton’s desk model, it will be both novel and historical, as it will be the first working “Cray” in decades.

“The Cray-1 is one of those iconic machines that just makes you say ‘Now that’s a supercomputer!’” Fenton wrote on his blog in 2010. “Sure, your iPhone is 10X faster, and it’s completely useless to own one, but admit it … you really want one, don’t you?

35 Responses to “The search for the lost Cray supercomputer OS”

  1. Presumably you’re looking for COS, one of the operating systems that ran on the Cray-1 and the X-MP. However, the X-MP also ran UNICOS, a version of Unix, so it would still be historically accurate to just run Unix (or Linux) on your replica Cray.

    On the other hand, if you are actually emulating the Cray architecture then this could be difficult since Crays didn’t have virtual memory.

  2. I concur with the EPFL. I used to live near the EPFL and have seen this system. I would recommend contacting their CS Dept. Everyone at that school speaks English and/or French.

  3. Sacha Labourey

    Maybe, the most impressive of all of this, is that when he launches the console, you can see that the path to his project is “small_projects”. Maybe Andras Tantos can share some of his “big_projects” with us, they must be truly impressive given what the small ones already are :)

    Maybe the still has some resources since they had multiple generations of Cray (some still exposed in the corridors) and also a Museum, some links:
    – some history:
    – one of the machine in the corridors:
    – an article (header in English, content in French) on some of the Cray and other machines at EPFL:

  4. Hmmmm…actually, I think I just remembered the name of the guy that “owned” CSIM in 1985 when I was using it: Ben Johnson. There are probably only 1,000 Ben Johnsons in MN!

    Ben was the maintainer of CSIM when I was in COS testing, but not the original author (but Ben would remember who that was, I’m sure, if he sees this article).


  5. (I tried entering this and it didn’t appear to take – sorry if this is a duplicate post).

    One of my first projects at Cray (I was there from ’85-’95) was to test parts of COS that Don Lee wrote. He was kind of an eccentric character.

    The biggest challenge of doing what these guys are trying to do is that all of that code was written in CAL (Cray Assembly Language), and even if you had a CAL assembler available to you these days, it would be generating binary for a Cray-1/XMP CPU, not some off-the-shelf commodity processor like these guys are using. That’s probably what they mean when they say they created a simulator.

    Some of my early COS Testing projects ran in CSIM (Cray SIMulator). CSIM was written in FORTRAN, so if one could lay their hands on the CSIM source code, you could probably get it to compile in a “modern” FORTRAN compiler after some tweaking, and then you would have a VM that would run Cray binary unchanged. Of course, there could have been parts of CSIM that were written in CAL, so that would require a wholesale rewrite.

    I don’t recall the name of the guy that wrote CSIM, but that would be the guy for these folks to talk to. Maybe he still has the code on a DG disk pack somewhere. :)

    Good luck!

    • Both Chris’s FPGA Cray-1 and Andras’s simulator are running Cray binary code. The FPGA only replicates the vector processor. The simulator covers the entire gamut of devices that were required to boot, run, and interact with an X-MP: the IOP, disk drives, terminals, and more. The simulator is able to boot from and run the recovered COS image; this obviously wouldn’t work if he weren’t working with Cray binary code. Download it and give it a try!

  6. Jon Bringhurst

    We still have our Cray-1 (with serial number #001). It’s sitting behind a pane of glass in the Bradbury. If anyone has a serious proposal for novel research, there are people here who would be willing to consider it.

  7. Loren Mastin

    I worked at Cray and I know a programmer who was there from the beginning who worked on the OS. I can give you his name but he moved away several years ago and we’ve lost contact. This is awful – Paul Johnson – age @60, once lived on Moonlight Drive in Chippewa Falls WI.

    Also, we have a supercomputer museum in Chippewa Falls:

    Chippewa Falls Museum
    21 E Grand Ave, Chippewa Falls · (715) 720-9206

    I hope this helps some.

  8. Jack N Fran Farrell

    Too bad there aren’t a lot of Cold War Surplus warehouses full of computers like the WWII aircraft and hangar queens. Clubs could relive the computers that overwhelmed the USSR by running them seven decades after the wars end.

  9. I worked in a lab that used a small VAX to preprocess data to prepare it to feed to a Cray, and interpret the result from the Cray run. We got so many minutes of Cray time per week, so we would gather data, input it on the VAX, and run a validation job to make sure the data was reasonable. My job was to run the VAX job to package the data and command set as a batch job for the Cray, transmit the batch job to the supercomputer to wait for our turn in the queue, and then print graphs based on the data returned. If I remember correctly, in less than 30 minutes, the Cray ran fourier transform on a large enough data set that our little VAX couldn’t have managed it. Certainly not in the week we had between data dumps.

  10. According to some Cray employees who were around when SGI sold Cray to Tera, the root of the problem is that SGI ordered the destruction of all legacy Cray software before the spin-off.

  11. I disagree with the “as it will be the first working “Cray” in decades” point. There are many working Crays in the world, in fact Cray (in one form or another) still manufacture machines today.

    Even if you said “as it will be the first working Cray 1 in decades” I would still disagree as I heard that there was still one running in an Arab state doing something with oil exploration as late as 2008.

  12. Jon Pugh

    We had Cray-1 #6 at NMFECC, later renamed NERSC, but we ran our own OS on it written in FORTRAN. That was called CTSS for Cray Time Sharing System and was for many years the only interactive OS available for the Cray. We also had one of the first Cray-2 machines, a four processor unit which was able to perform over a gigaflop.

    NERSC still exists at LBL, but I don’t know if they’ve saved any of that early software. Even if they did, getting it read in, compiled, running and debugged would be quite a chore.

    • Alex Kluge

      I cut my teeth on large scale & parallel computing on these systems. First remotely from the Princeton Plasma Physics Lab, then directly from LLNL. This is the time frame where the foundations of parallel and cloud computing were laid down. The high end processing was dominated by Cray.

  13. Kcuhc Ttenneb

    When I was in college (81-85)I, used to babble on to my friends about just such a thing. I even used to chant “Desktop Cray is On It’s Way” to annoy them because they thought it was so preposterous. I guess I’ll have to change it to “Desktop Cray is Here to Stay!” Thanks, guys!

  14. The crazy thing is there are scores of systems that government, rail and heavy industry rely on where the original software is not being preserved and cannot be replicated.