Sometimes it is hard to remember just how far computers have come in the last three or four decades. An old NASA video (see below) has been restored with better sound and video recently that shows what passed for a giant computer in 1986. The Cray 2 runs at 250 MHz and had two gigabytes of memory (256 megabytes of million 64-bit words).
Despite the breathless praise, history hasn’t been kind to the Cray 2. Based on ECL, it had 4 processors and –in theory — could reach 1,900 megaFLOPs/second (a FLOP is a floating point operation). However, practical problems made it difficult to get to that theoretical maximum.
Still, it was the fastest computer of its day and that amount of memory in 1986 was stunning. A 1987 Radio Shack catalog, for example, shows a Tandy 1000 SX that could run at over 7 MHz, had 384K of RAM, and not one but two floppy drives for only $1,199.00. Well, you’d need to pay extra for the monitor, but still. The year after that, $2,600 would get you the 4000 that ran at 16 Mhz with an 80386 onboard. For that price you got a full megabyte of memory. There were two empty slots for a hard drive and you’d still need to buy both a monitor and a display card! The Cray had 2,000 times that much memory and could probably get to it a lot faster, too. Of course, it also cost around $16 million. It is always dicey to compare speeds of different computers, but we read that the iPad 2 is about as fast as the Cray 2.
Of course, the sticker price was just the start. The 5,500 pound computer could burn up to 200 kilowatts of electricity and we have a feeling the special 3M liquid that cooled the thing wasn’t that cheap either. Apparently, some lasers use a similar fluid and we found some of that currently available for about $500 for a 1000 cc container.
Computer History has a nice scan of an old Cray sales brochure you might enjoy. We got a chuckle out of the page describing it as “compact” with 14 vertical columns taking up “only” 16 square feet of floor space. You had to have space for the 320 pluggable modules. For a sales brochure, it goes into a lot of detail with block diagrams and discussions about the actual architecture. The memory bandwidth, according to the brochure, is a cool 8 gigabytes per second. We also found it interesting that the system had motor generators to produce the 400 Hz power the system needed.
If you don’t have 16 square feet and the three-phase power service to operate one of these, we’ve seen a knock off Cray-1 that looks cool. Or you can disguise your PC, although if it isn’t water-cooled, that seems somehow inappropriate.
I saw a decommissioned one in a computing center.
It was surprising because it was smaller than its other decommissioned family members…
That was (and still is) Cray’s schtick: packing as much performance in as little space as possible.
In the case of the Cray-2, the density was also needed to keep signal paths short for fast cycles. This was a problem since it was built on current-sucking ECL. You can’t put heatsinks on tens of thousands of ICs, hence the direct immersion cooling.
The power supplies, I/O, and the spectacular Fluorinert waterfall heat exchanger were in separate cabinets, so what’s left is a surprisingly small chassis.
“The 5,500 pound computer could burn up to 200 kilowatts of electricity and we have a feeling the special 3M liquid that cooled the thing wasn’t that cheap either.”
I imagine computer museums must have a big bill.
“We got a chuckle out of the page describing it as “compact” with 14 vertical columns taking up “only” 16 square feet of floor space. You had to have space for the 320 pluggable modules.”
Kaypro compact.
Here is a quote for one type of fluoroinert…
http://synquestlabs.com/product/id/52530.html
And if you’d like to purchase a 750 lb (341 kg) drum of another type of fluoroinert…
https://www.chemworld.com/Novec-FC-770-Fluorinert-Electronic-Liquid-p/novec-fc-770.htm
It’s a little spendy because they have to milk it from this single species of wild gnat that lives by a glacier fed mountain stream high in the swiss alps, and there’s only a 2 month season.
Improbabilimide of unobtanium would do in a pinch. Or unicorn tears.
“8 billion gigabytes per second” sure is pretty fast.
I checked, the linked PDF says Total memory bandwidth is 64 gigabits or 1 billion words per second.
Fast for back then. Modern GPUs are in the hundreds of GB/sec for memory bandwidth.
I used to work at a place in the early 90’s that had a Cray. Can’t remember the model, but I remember it was red. I even had an account on it! They bought it used. I seem to remember the coolant and power ended up being about the same cost as the computer after 2-3 years. Later, after I left that job, someone told me it became an ornament in another companies lobby.
Yes, I’m sure multiple people have spotted the error and let it go as a typo or miscalculation. I know I enjoyed the write-up and the nostalgia of seeing a video about an old machine from the 80’s doing state of the art things for the time. Making multiple posts pointing out the mistake doesn’t really add anything in my opinion. It seems a bit on the pedantic side. I do agree with your last sentence.
When I see people outraged over some writer’s minor typo and calling out them out by name I always think – huh… I guess the writer slept with his wife or something.
Honestly, I knew what was meant. And at the end of the day didn’t do me any harm. I’m not basing business decisions or investing my life savings based on a historical article.
The interweb is grand. I wonder how many people would be that boorish if they met some of these network writers in person?
You must be beyond pathetic as a person.
Cray were sold in any color the owner wanted. If someone’s going to spend several million dollars for one computer, it was no problem for them to custom paint any of the exterior parts. I recall one Cray even got leather covering.
8 billion bytes is what I meant, of course. Just got carried away on prefixes.
reminds me of this Neil Cicierega song:
https://www.youtube.com/watch?v=AhSEfKo0tlw
I got to see one earning it’s keep in the mid 90s, not real close up though, it was just lookenpeeping at the blinkenlights. Beowulf systems were just getting traction and beginning to outperform these so can’t imagine they kept it going much longer.
Me too. As a kid, at NCAR in Boulder, they had a sweet Cray (running?) on display. Dang sexy in the mid 80’s.
Yeah, the last I saw (Serial # 3?) was used as a bench outside the main machine room, but that was 20+ years ago…
Hmm. If I’m going to pay 16M for a computer, it better come with seats.
https://en.wikipedia.org/wiki/Cray-1#/media/File:Cray-1_(1).jpg
How cool would it be if it were cooled with butter, from which the heat were removed in turn by expanding the water content of dried corn kennel into steam?
Of course if it’s going to make popcorn, you’d want something to actually watch. Perhaps the entire beta-max catalog?
Arrrgh, kernels not kennel. Bother ..
You could probably fit the whole Betamax catalog on a phone now using compression designed to trade space for CPU power, and then have the equivalent processing power of this available to decode those streams in real-time. Gotta love Moore’s Law.
We went from using parallel processing being the way to handle iteration over large data sets, to now parallel processing being used just to keep Moore’s Law alive.
All of this has happened before. All of this will happen again..
But paralllel processing hasn’t got anything to do with Moore’s law, as it doesn’t measure system performance.
Moore’s Law gives you transistors, but communicating synchronously across a huge plain of transistors costs a ton in space for traces and also power, so you make a bunch of smaller localized units and do your processing in parallel. I think it’s a justified statement.
Moore’s law only deals with the economics of transistorized chips. It hasn’t got anything to do with parallel processing, processing performance, or data handling capacity.
Moore’s law is simply the number of transistors on a microchip at the most economical (lowest cost) chip size. It’s a function of wafer size, processing technology, and process yield.
“two gigabytes of memory (256 megabytes of 64-bit words).”
Um….okay.
Want to take a deep breath and restate that?
I’m sure that was an honest typo and I guess fixed, but actually before the ISO standard in (i think) 1993 a byte wasn’t always 8 bits. It was usually the smallest word size the computer could handle. Don’t know if that’s 64 bits or not on a cray, but it was perfectly correct to say machine X has 100 megabytes of 12 bit bytes. Until the standard. That is why a lot of specs say octet instead of byte.
Come on. A byte has been 8 bits (or an ASCII character) since the stone age. Check the cave paintings and Runipedia.
I have a bunch of different byte-width systems sitting in my office, from some 6-bit embedded chips up to some complex 28-bit control systems. Bit-width is up to the designer of the system and almost always what is most efficient for the task the system was designed for.
One of the systems I am working on a computer that was built to control a bunch of lab equipment in real-time based on a set of 12-bit ADCs, so the whole thing was designed with 13-bit memory locations, 13/26-bit buses. It doesn’t display much in the way of text, so the wasted bits when displaying text are trivial when compared to the savings on the calculations.
“a FLOP is a floating point operation” while that is literally true, that’s not how that particular unit of measure works. It’s actually FLOPS, or FLoating-point OPerations per Second, therefore it is both singular and plural; thus, that Cray 2 performed at 1,900 megaFLOPS, not megaFLOPs/second (which is redundant; that would make it millions of floating-point operations per second per second). The Summit mentioned in an earlier comment operates at 1.88 exaops, or 1.88 billion billion calculations per second using a mixture of numerical precisions, or apparently 122.3 petaFLOPS when operating in a typical floating-point environment.
When we got a Cray at my first job, my colleague Nero reckoned FLOP stood for a “f****** lot of processing”:
http://www.chilton-computing.org.uk/gallery/ral86/orig/r86r5865.jpg
B^)
Matt Brunton…. Simon here, I was just recalling the cooling system failure alarm I made for that beauty!
Hope you’re fairing well, we really ought to have another Christmas Reunion And Party before we all get too old to remember!
Though you show that to a computational fluid dynamicist now and they’ll cringe like you showed them cave drawings. Or you say something like, “Hey I wish I could run that particular earlier type of analysis on my desktop to figure out an antenna housing for my car, because after all that’s how they designed the fricking space shuttle…” and you’ll get yelled at and told you absolutely need to rent time on a thousand xeon box to do such a thing or you’ll blow your car up.
Ooops, somehow ended up here, meant to be in reference to last vid on main post.
We’ll strictly speaking when you first go from 0 to 1,900 megaFLOPS you are also operating at 1,900 mFLOPs/second over the first second :)
That’s why it is a small s… and I added /s — just for clarity, but yes.
Interesting… 1986, 2 gigs of memory and 2 GFLOPS, $16 million. 2001, 2 gigs of memory and 3 GFLOPS (dual Alpha 1U server), $9000. 2018, 1 gig and
Oops. 2018, 1 gig and about 4 GFLOPS (Raspberry Pi), $35.
That is interesting.
2018 Now that every researcher has computing power beyond their wildest dreams and can solves the greatest of problems, we have porn delivery rates that boggle the mind!
And all our software is scripts running in a virtual machine on 12 layers of abstraction, which takes up 99% of the processing resources.
There’s one of them Crays at EPFL in one of the hallways. Heard they use it as a bar on some events there.
I worked on getting TCP/IP running over Hyperchannel on that machine at EPFL in 1988.
Wow, thanks for the trip through the wayback machine!
There is also a cray 2 at the Heinz Nixdorf museum in Paderborn, for those of you who are in Germany.
Way back when, Seymour Cray Worked for Control Data Corporation and had designed the 6000 series computers for them. I had worked for CDC at various aerospace companies back in the mid 70’s. as an engineer on those beasties and also the more advanced 7000 series. Those machines were simply cooled with Freon running through evaporator coils in the aluminum backplane of the chassis. critical timing between the cordwood modules which contained the actual circuitry was by way of wire length. The wire length determined the amount of delay in nano secs. between the various registers and channels. Great pains went into getting these just right. And many countless hours were spent in tracing faulty circuitry.
I believe having the wires be the right lengths was why those initial Cray units were circular. More volume on the outer diameter for cooling, less distance on the inner diameter for signals to travel, and you could fit a human in there to run the wiring.
The motor-generator sets are really not a bad idea. The 400Hz AC can use smaller transformers which are easier to position wherever they’re needed inside the unit, and the rotating mass of the motor-generator system serves as a line stabilizer and a crude sort of very-short-term UPS. Any additional flywheel would serve to increase ride-through capability.
Such a system is a “brick wall” for surges and transients, as the motor simply can’t accelerate that fast, and it would likely fly apart before the 400Hz side would’ve reached a voltage that could damage the computer.
They can also be used for galvanic isolation, if the shaft connecting the motor to the generator isn’t electrically conductive. This system is used, with fiberglass shafts, in the high-voltage ion generators that feed certain particle accelerators.
Ah, the motor-gens. These were housed down in the basement, at least at the Aerospace Corp. Gee, I believe that there were three separate units there. But not really much of a ups. At the first sign of any failure, all power circuits would go open. Immediate termination of power to everything on the floor, including peripherals. I do agree with them blocking any transient voltage spikes.
In the Cray-2 the smallest addressable memory unit was 64-bits. This means that the byte size was 64-bits therefore it had 256 million 64-bit bytes. The Cary-2 was did not have a byte oriented architecture.
All the Cray machines has an 8-bit byte when programmed in C.
You are right about 64 bits being the minimum memory operation though.
Without 8-Bit bytes, I don’t think we would have ever gotten networking working from C. The UPPER three bits in the address were used to hold the “byte within the word” information. ????
The structure packing and unpacking was a real pain however.
Look at the differences between the X10 and X11 protocols to see how packing on the Cray was handled.
I don’t want to go back to those days.
I should have said all the CRAY RESEARCH machines had an 8-Bit byte. Sorry.
I don’t care – I still want one. Maybe power it up on special occasions…
I have the same feelings. I wish I had not tossed out my Commodore 64!