Join 3,416 readers in helping fund MetaFilter (Hide)


Homebrew Cray-1A
August 31, 2010 12:52 PM   Subscribe

Lots of people remember the Cray-1A. And how could you not, it was a supercomputer with built-in seats. But Chris Fenton has done more than just reminisce, he he built his own 1/10 scale Cray-1A. It's not just a model, it actually runs Cray software. The only problem is there doesn't seem to be much Cray software on the net, so if you have some let him know.
posted by tommasz (66 comments total) 12 users marked this as a favorite

 
Have we killed his website already?
posted by PeterMcDermott at 12:57 PM on August 31, 2010


80 megaflops, eh? So, a Playstation 3, more or less?
posted by Joe Beese at 1:00 PM on August 31, 2010


Note To Self: Do not attempt to use 1980s era supercomputer as webserver.
posted by The Bellman at 1:01 PM on August 31, 2010 [5 favorites]


Must be running his site off of a 1/10 scale TI99/4A
posted by mr_crash_davis mark II: Jazz Odyssey at 1:02 PM on August 31, 2010 [3 favorites]


Also: the vintage calculator museum is a few threads down.
posted by Joe Beese at 1:03 PM on August 31, 2010


But can it run Crysis?
Heh.. Ahem.
posted by pyrex at 1:05 PM on August 31, 2010 [1 favorite]


I have a shirt that says "My other computer is a Cray" with an Apple logo...I'm hoping to wear it one day and someone cool will recognize it, this has happened only once.
posted by hellojed at 1:06 PM on August 31, 2010


It's pretty astonishing that what weight five and a half tons in 1976 can be duplicated thirty-some years later by four cell phones running in parallel. (iPhone benchmarks at 36MFLOPs, Cray-1 generally hit around 136).
posted by CheeseDigestsAll at 1:07 PM on August 31, 2010 [3 favorites]


In case you're interested, PS3 does 218 GFLOPS (CRAY-1 could do 0.250 GFLOPS)
posted by East Manitoba Regional Junior Kabaddi Champion '94 at 1:09 PM on August 31, 2010 [3 favorites]


Hmmm, the link worked fine for me at first but subsequent visits haven't. Should this post be deleted until the site becomes stable? I'd be fine with that.
posted by tommasz at 1:09 PM on August 31, 2010


Today's fastest (unclassified) computer is also a Cray.
posted by East Manitoba Regional Junior Kabaddi Champion '94 at 1:11 PM on August 31, 2010


NYUD Mirror of the OP is working for me.
posted by Skorgu at 1:15 PM on August 31, 2010


Have we killed his website already?

No, Slashdot did, all of 3 hours ago now.
posted by splice at 1:16 PM on August 31, 2010 [1 favorite]


There is even more to learn about Cray!
posted by TwelveTwo at 1:16 PM on August 31, 2010


Cool project but I'm really surprised that zero Cray software survived. That's a real tragedy. Certainly a quality FORTRAN compiler is about the only thing that makes a highly pipelined vector machine like this at all usable.
posted by GuyZero at 1:18 PM on August 31, 2010


It's pretty astonishing that what weight five and a half tons in 1976 can be duplicated thirty-some years later by four cell phones running in parallel. (iPhone benchmarks at 36MFLOPs, Cray-1 generally hit around 136).

In fairness, the Cray-1 had seats. And though you can sit on an iPhone, it voids the warranty.
posted by Joe Beese at 1:19 PM on August 31, 2010 [11 favorites]


All of my programming so far has just been in straight octal machine-code, just assembling it in my head.

Um... wow.
posted by rkent at 1:20 PM on August 31, 2010


My dad worked in the Supercomputer division at IBM in the 70s and early 80s. He was always going off about 'those bastards at Cray were in there talking to my customer blah blah'. So whenever I hear Cray all I hear is him on the phone at night bitching to his team.
posted by spicynuts at 1:20 PM on August 31, 2010 [3 favorites]


Man. Have to say that Crays look fucking awesome, even all these years later.
posted by Threeway Handshake at 1:22 PM on August 31, 2010 [1 favorite]


In fairness, the Cray-1 had seats. And though you can sit on an iPhone, it voids the warranty.

But then, so does running the Cray software.
posted by Pogo_Fuzzybutt at 1:24 PM on August 31, 2010 [5 favorites]


I have a Cray mug, which my father-in-law gave me after I basically nerded out over it. I'm pretty sure he has no idea of it's significance. I love that mug.
posted by Artw at 1:27 PM on August 31, 2010 [1 favorite]


Thanks, Skorgu, for the NYUD link. they don't allow us to access cache/mirror sites here at work.
posted by tommasz at 1:33 PM on August 31, 2010


I love that mug.

The only Cray mug on ebay is going for $75!!!!
posted by GuyZero at 1:36 PM on August 31, 2010 [1 favorite]


The CRAY-1A had... 1 million 64-bit words of high-speed memory.

I picture the operators lording it over their less privileged colleagues...

That's right, losers. 8 megabytes of RAM!
posted by Joe Beese at 1:36 PM on August 31, 2010 [3 favorites]


At the time the Cray 1A came out most computers had 10's of kilobytes of memory. It's nearly three orders of magnitude (800x) more memory than an average computer of the time.

In comparison, if we assume an average computer today has 4 GB of memory, it's like having 3 TB of main memory. I don't know if there are non-supercomputers with more than 1 TB of RAM or if any computer has that much main memory.
posted by GuyZero at 1:43 PM on August 31, 2010


Since design mattered back in the day, a favorite was the Thinking Machines CM-5 with design by Maya Lin. Looked like a lightning bolt from overhead, and more blinking lights than you really needed to do the science. But it looked like a computer. When retired, the Cray Y-MP supposedly was eclipsed in computational power by the Nintendo 64. Rumor was that they would give it to anyone who would pay to have it removed from the building (required a crane).

Back when supercomputers were really super expensive the strategy was to drop them in high performance computing centers so that using that Internet thingy research scientists could have access to some serious power. Al and Tipper Gore paid a visit on the 88 campaign and WE at least understood why he was important.

I used to do my statistics homework running SAS on the Convex, just because I could.
posted by cgk at 1:45 PM on August 31, 2010 [3 favorites]


In comparison, if we assume an average computer today has 4 GB of memory, it's like having 3 TB of main memory. I don't know if there are non-supercomputers with more than 1 TB of RAM or if any computer has that much main memory.
The motherboard I have right now supports 128 GB of ram, which is 1/8th of a terabyte. It would only take 24 of them to get up to 3TB. Most 'supercomputers' these days are just lots of PC motherboards networked together, and there are definitely clusters with 3TB of memory.

(Especially when you consider that regular flash memory is probably faster then the RAM in a cray, and you could easily build a machine with 3TB of flash)
posted by delmoi at 1:47 PM on August 31, 2010


I don't know if there are non-supercomputers with more than 1 TB of RAM or if any computer has that much main memory.

Windows 7 maxes out at 192 GB, apparently. But perhaps it could be done in Linux?
posted by Joe Beese at 1:49 PM on August 31, 2010


I don't know if there are non-supercomputers with more than 1 TB of RAM or if any computer has that much main memory.

Yeah, you can get single servers with 1TB of RAM. Example.

Off the shelf desktops not so much, as delmoi points out.

Windows 7 maxes out at 192 GB, apparently. But perhaps it could be done in Linux?

Definitely could be done in 64-bit Linux. A 64-bit Windows server OS (e.g., Server 2008 R2) could handle it as well.
posted by jedicus at 2:00 PM on August 31, 2010


Some interesting comments in that Slashdot thread.

The seats were uncomfortable!
posted by Joe Beese at 2:07 PM on August 31, 2010


Most Dell server boxes with the standard intel architecture max out at 1 TB of RAM. The IBM x3850, a favourite of virtualized datacenter admins, also tops out at 1 TB. Although after a bit of poking around HP's Integrity NonStop NS14200 Server supports 16 TB RAM and the NS16200 supports 127 TB but I don't think these are the same kinds of systems as a standard server and are more likely blade systems, I'm not sure.

Most 'supercomputers' these days are just lots of PC motherboards networked together, and there are definitely clusters with 3TB of memory.

So to make the Cray seem more impressive I'm ignoring clusters. Certainly I could run a job right now and allocate 3 TB of memory to it (ok, not really, i don't have that kind of machine quota) but that would be across a lot of machines. Having a monolithic architecture is an entirely different beast vs a distributed one.
posted by GuyZero at 2:08 PM on August 31, 2010


They also informed me that apparently SGI destroyed Cray’s old software archives before spinning them off again in the late 90’s.

If that's true ... fuck SGI.
posted by Kadin2048 at 2:08 PM on August 31, 2010


I installed a T1 to the spinoff that still was vaguely related to Cray in Chippewa Falls, WI (like the bits of Cray that didn't pack up and go out West) sometime in '97 or '98 or so. They had a Cray 1 in the lobby, theoretically as seats for visitors, but I don't think anyone ever sat on the bench. Too reverential.
posted by Kyol at 2:09 PM on August 31, 2010


Mac OS X 10.6 evidently can handle 16 terabytes of RAM.
posted by Scoo at 2:15 PM on August 31, 2010


Sigh, I had a manual for one of these things. It got lost during the great move along with 2 plastic cases of PDP-11 diagnostics on punch tape.
posted by zengargoyle at 2:23 PM on August 31, 2010


Heh, no it can handle 16 EXABYTES of RAM or 16 million terabytes.

And you thought the existing Mac Pro machines were expensive.
posted by GuyZero at 2:26 PM on August 31, 2010 [1 favorite]


Mac OS X 10.6 evidently can handle 16 terabytes of RAM.

Re-read that link. Mac OS X can theoretically handle 16 exabytes of RAM (16 billion GB), but actual Mac Pros can only be equipped with 32GB today. Xserve servers can be equipped with 48GB (possibly quite a bit more if you put in 8GB or 16GB sticks).
posted by jedicus at 2:26 PM on August 31, 2010


When Cray-1A came out, it was described in an April edition of some computer magazine. Traditionally, the computer magazines then would put in a lot of April Fools jokes in their April editions. I thought the description was just another one of those jokes, because it was so outlandishly impossible. I still think it is one of the coolest looking computers ever built.
posted by Xoc at 2:32 PM on August 31, 2010


My father-in-law used to work at Cray. Doubt he has any old software laying around, but, I forwarded it on to him.
posted by trbrts at 2:59 PM on August 31, 2010


Some ubernerd must have a Cray version of Doom somewhere...
posted by starman at 3:05 PM on August 31, 2010


I'm sure if I forage through some of the disused offices here I can come up with ex-Cray memorabilia. I know at least one person who's retirement plan is selling Cray stuff, including components, on eBay. One of my co-workers has some of Seymour's original design documents.
posted by substrate at 3:08 PM on August 31, 2010


Mac OS X can theoretically handle 16 exabytes of RAM (16 billion GB)

Yeah, if Apple went to that magical hardware platform where all 64 bits were used for addressing.

Rumor was that they would give it to anyone who would pay to have it removed from the building (required a crane).

We currently have a bunch of old Sonus equipment taking up space in our racks because not only will nobody buy them, it's going to cost us money for somebody to take them off our hands.
posted by kmz at 3:08 PM on August 31, 2010 [1 favorite]


All of my programming so far has just been in straight octal machine-code, just assembling it in my head.


It's statements like that one that make me immediately picture him waking up in the morning, and speaking the Mentat chant:

"It is by will alone I set my mind in motion.
It is by the juice of Safu that thoughts acquire speed,
the lips acquire stains the stains become a warning.
It is by will alone I set my mind in motion."
posted by chambers at 3:12 PM on August 31, 2010 [3 favorites]


How did i get so much Dune in my metafilter all the sudden?
posted by clockwork at 3:18 PM on August 31, 2010


Yeah, if Apple went to that magical hardware platform where all 64 bits were used for addressing.

It's true that right now x86-64 systems as implemented only support 48-bit addressing (allowing for a mere 256TB of RAM), but the architecture supports 52-bit addressing for physical memory and 64-bit addressing for virtual memory.
posted by jedicus at 3:25 PM on August 31, 2010


It's pretty astonishing that what weight five and a half tons in 1976 can be duplicated thirty-some years later by four cell phones running in parallel. (iPhone benchmarks at 36MFLOPs, Cray-1 generally hit around 136).

...and yet neither support Flash!

Do you prefer your closed system gated or seated?
posted by mazola at 3:35 PM on August 31, 2010 [3 favorites]


The iPhone will never hit a running throughput of one operation per CPU clock cycle, so in that sense it will never be fully duplicated by most modern systems. Comparing a 1GHz processor to 80 Mhz is cheating by throwing more cycles at the problem.
posted by GuyZero at 3:39 PM on August 31, 2010 [1 favorite]


The iPhone will never hit a running throughput of one operation per CPU clock cycle

I don't know that much about the ARM processor used in the iPhone, but I know Intel CPUs since the Pentium have done pipelineing and instruction level parallelism that gets pretty high throughput.

Combined with lots of floating point hardware, and SIMD you can get really high floating point performance. Looking at the wikipedia page on ARM (scroll down to the table) it looks like the ARM cortex A8 can actually handle about two integer ops per clock.
posted by delmoi at 4:07 PM on August 31, 2010 [1 favorite]


"Combined with lots of floating point hardware, and SIMD you can get really high floating point performance. Looking at the wikipedia page on ARM (scroll down to the table) it looks like the ARM cortex A8 can actually handle about two integer ops per clock,"

...as the actress said to the Bishop.
posted by Sebmojo at 6:43 PM on August 31, 2010


Combined with lots of floating point hardware, and SIMD you can get really high floating point performance. Looking at the wikipedia page on ARM (scroll down to the table) it looks like the ARM cortex A8 can actually handle about two integer ops per clock.

Yeah, I heard cortex was totally SIMD and could support up to 5 ops per clock. But, then, he's got the mod panel.
posted by Netzapper at 7:11 PM on August 31, 2010


The Sun M9000 can handle 256 cores and 4 TB of main memory. It seems most of the current builtto order windows boxes top out at 128-192 MB.
posted by stp123 at 7:15 PM on August 31, 2010


Amdahl's Law, people. MPP only gets you so far. In our wintel-world, we like to convince ourselves that Moore's Law is working for us, but in reality, it's working against us. We're using that transistor density to cram more of the same old shit on a shiny new shingle rather than finding new and radical new uses for it.

I'll be blunt - modern hardware bores me. Video cards with amazing chips that live in a perpetual second-class status and have their own RAM is exciting... and it's used for nothing, absolutely nothing, other than games. Maybe some animation rendering. Dull, old crap.

I need a new Amiga, something cheap and powerful beyond the sum of its parts. I need a new Mac, an interface for a modern, smartphone-tablet-desktop-workstation-server-supercomputing world, a fuzzy cloud that can adapt to what's needed when.

I'll also need a new programming paradigm to allow the everyman to access this high-end kit for their own needs, easily and seamlessly. (Cuz, dude, when a non-geek asks you how to program their computer, and your first instinct is to tell them to develop test cases, something is terribly wrong.)
posted by Slap*Happy at 7:51 PM on August 31, 2010 [2 favorites]


x86_64, Linux, 2048 cores, 16 TB RAM, in a single system image.
posted by theclaw at 7:58 PM on August 31, 2010 [3 favorites]


... that's shipping today.

Slap*Happy, There's a lot happening with GPU computing.
posted by theclaw at 8:02 PM on August 31, 2010


Video cards with amazing chips that live in a perpetual second-class status and have their own RAM is exciting... and it's used for nothing, absolutely nothing

Yeah, new/future builds for most real web browsers are starting to use the GPU to render pages.
posted by Threeway Handshake at 8:16 PM on August 31, 2010


"Gregory Volfovich Chudnovsky recently built a supercomputer in his apartment from mail-order parts." The Mountains of Pi by Richard Preston in The New Yorker, March 2, 1992.
posted by neuron at 8:31 PM on August 31, 2010


I'll be blunt - modern hardware bores me. Video cards with amazing chips that live in a perpetual second-class status and have their own RAM is exciting... and it's used for nothing, absolutely nothing, other than games. Maybe some animation rendering. Dull, old crap.

What?

First, as theclaw mentions, there's CUDA. There's a C compiler targeting the GPU, with a nearly-standard C language implementation. Obviously the paradigm is radically different, but it's pretty directly accessible to anybody smart enough to recognize data parallelization opportunities.

I see absolutely no reason that the GPU shouldn't sit on the PCI-e bus and perform data-parallel operations on a second-class, as-needed basis. To get the kind of compute that makes a GPU shine, it's gotta be pretty application-specific anyway. So, abstracting it off as a transparent OS service seems sub-optimal. And you can't make vital OS services themselves dependent on having a GPU.

Second, operating systems already draw the UI with the GPU. All three operating systems use GPU-accelerated GUIs when a GPU is available. On linux and mac, this translates to real productivity improvements, as features like Expose and window previews become viable. For instance, I have Meta+scrollwheel mapped to a magnifying glass. I can zoom in to about 12x before shit starts getting blurry, which makes it a helluva lot easier to get a good look at the [CENSORED] of the girls in the pictures.
posted by Netzapper at 8:51 PM on August 31, 2010


Good find on the Altix with 16TB memory. The Sun M9000 seems less impressive than I remember Sun boxes being at the top end but it's still very, very big. The IBM zEnterprise mainframe goes up to 3TB main memory.

At any rate, it does show how much the world of computing has changed at the high end. No one really even tries to build these massive monolithic systems any more except for enterprises who have huge systems that can't be re-architected. And the occasional scientist doing simulations that require a single memory image or that are limited by inter-processor communication bandwidth.

But big systems like this have really been eclipsed. I maintain that even a lowly Cray 1 has technical achievements that modern systems can't touch, but, of course, they turned out to be things that ultimately do not scale as well as basic clock increases and the fundamental shift to distributed systems.
posted by GuyZero at 9:01 PM on August 31, 2010


I'll also need a new programming paradigm to allow the everyman to access this high-end kit for their own needs, easily and seamlessly. (Cuz, dude, when a non-geek asks you how to program their computer, and your first instinct is to tell them to develop test cases, something is terribly wrong.)

Making the intractable tractable is a project that humankind has been working on for thousands of years.

The majority of programming work that is done when creating human interfacing systems is talking to stakeholders and codifying their desires, budget, and future capabilities to discrete logical structures and algorithms.

When the everyman is capable of translating their thoughts to a logically consistent form, we can start making all those boxes for you to move around with your mouse.
posted by sleslie at 11:25 PM on August 31, 2010 [2 favorites]


GuyZero beat me to it. The IBM zEnterprise 196 - the latest IBM mainframe - is impressive stuff, see here and here and also my blog here with more onward links. It also runs Linux, and they say one box (the size of a large fridge) could host up to 100,000 virtual Linux servers.
posted by magpie68 at 2:04 AM on September 1, 2010


Yeah, new/future builds for most real web browsers are starting to use the GPU to render pages.

Windows has been able to do that since Vista. If you turn on desktop composition, your whole desktop is rendered in the GPU.

So, I wish they wouldn't do that. The OS should be handling that stuff.
posted by gjc at 3:44 AM on September 1, 2010


Today's fastest (unclassified) computer is also a Cray.

These super-fast computers on the TOP500 list -- what are they used for?

Also, how do these supercomputers differ from large clusters of non-super computers? Are they really different from Google's datacenters full of normal boxen all working on a common application?
posted by Galaxor Nebulon at 10:13 AM on September 1, 2010


These super-fast computers on the TOP500 list -- what are they used for?

Simulations, mainly. The prototypical example is weather simulation. But, also protein folding and other molecular biology. Also computational chemistry. And crunching numbers from high-energy physics experiments.

These are all tasks where a huge problem can be broken down into small pieces, each piece can be computed independently, and then the pieces can be rearranged into a complete whole.

Also, how do these supercomputers differ from large clusters of non-super computers? Are they really different from Google's datacenters full of normal boxen all working on a common application?

In two ways:

The first is performance. The nodes in a supercomputer are interconnected in such a way that they can communicate with each other much better than standard data center machines. They also often have architectures poorly adapted to running general-purpose programs, but that are freakin' awesome for the sorts of data-parallel problems I mention above.

The second way that they differ is in programming. In general, the supercomputers present themselves to a programmer as a single computer. So, there's rarely a need to manually assign work units to nodes, or to explicitly move data from one place to another. These machines are built in such a way that the heavy lifting of managing thousands of compute nodes is largely taken care of by the operating system.
posted by Netzapper at 10:20 AM on September 1, 2010


Most supercomputer are vector machines aka SIMD where they're optimized to do identical mathematical operations on huge amounts of data. Simulating nuclear explosions or tornadoes is just a lot of really big matrix multiplications. This is a very different problem from the map-reduce type operations that distributed databases (Google) do.

Essentially the US military stopped blowing up Nevada every few weeks to test nuclear weapons by buying supercomputers. The two biggest supercomputers are at Oak Ridge and Los Alamos which were historically two of the biggest nuclear research facilities on the planet. Now they do lots of stuff.
posted by GuyZero at 10:22 AM on September 1, 2010 [1 favorite]


gjc: If you turn on desktop composition, your whole desktop is rendered in the GPU.

No, this is manifestly incorrect. Composition and rendering are not the same thing. If you turn on desktop composition, each window is rendering to its own segment of memory; they may or may not use any GPU features to do so. Then the Desktop Window Manager acts as a compositor, and configures the GPU to do video scanout (i.e. telling the screen to draw stuff) of a single composite image that's created from the individual memory areas. There's not necessarily any hardware rendering going on whatsoever.
posted by atbash at 12:10 PM on September 1, 2010


In college I had access to lot of really cool computers. So one day, I was bored and went over to GaTech's GVU Lab and sat in front of an SGI Reality Engine 2. A beast of a graphics workstation that became famous for being used in Terminator 2. Being very bored, I telnetted to a Cray Y-MP I had access to. I think it was the Navy's. So, to pass the time I uploaded xtetris to the Cray, compiled it and run it and exported the display to the Reality Engine. Being single-threaded and not vectorized tetris run reallllly slow. So another student walks past me and he's like "why is tetris running like this on the SGI?". To which I had the pleasure if replying "oh that's just a DoD Cray".

Biggest waste of computing resources ever.
posted by costas at 12:23 PM on September 1, 2010 [3 favorites]


« Older Alan Jacobs laments the Hobbesian reality that is ...  |  A montage of movie dance scene... Newer »


This thread has been archived and is closed to new comments