6

I'm trying to find a quote that I'm 90% certain was attributed to Cray. It was to the effect that a computer that processes at a rate of one megaflop also needed a megaword of memory and [something else mega-sized but I can't remember what].

This has come up because someone has attributed an almost exactly identical quote to Amdahl, and now I'm wondering if I have a faulty memory, he does, or the entire collected world all has it wrong. And so, of course, now I can't find a reference to it at all!

3
  • 1
    I have heard a similar quote, and my guess would be Amdahl.
    – dave
    Commented May 11 at 0:15
  • 2
    The 'other one' is probably something to do with I/O bandwidth.
    – dave
    Commented May 11 at 0:23
  • 3
    Frank S. Preston, A Petaflops Era Computing Analysis. Report CR-1998-207652, National Aeronautics and Space Administration, Mar. 1998: "The wisdom of supercomputer architects in the 1980's was to recommend a memory size of one WORD per FLOP. (8 Bytes or 64 bits per word) Only the most well endowed facilities have been able to afford this is the past, and many operated with less than this .. . With the recent reduction in memory cost, the situation has improved. .. Current practice appears to size the memory in the order of one Byte per FLOP - a factor of eight less than the earlier goal."
    – njuffa
    Commented May 15 at 1:21

3 Answers 3

10

Well, I can at least tell you that it wasn't Amdahl. From "Gene Amdahl – A Personal Tribute," by John L. Gustafson:

Many computer architecture books and articles also point to Amdahl’s “guideline” that a well-balanced computer should have “a megaword per megaflop” or similar variants. I tried for years to track down the original source of this guideline, and could not find it anywhere. Then I realized I could simply ask him, since we were about to have lunch together. His answer surprised me: “I never said that.” Nor did he believe it was a good guideline! I suppose one of the hazards of being an industry icon is false attribution.

5
  • Nice find. The book being a great read for everyone interested.
    – Raffzahn
    Commented May 12 at 11:53
  • Well, there goes my best guess :-)
    – dave
    Commented May 12 at 13:00
  • Can we take someone's word that they never said something?  I can't remember things I said last week…
    – gidds
    Commented May 12 at 16:41
  • 2
    @gidds No, we can't. But this is direct evidence, that should be weighed along with that he thinks it's a bad idea (making it less likely he ever said it) and that nobody (so far) has been able to find an actual quote of him saying it.
    – cjs
    Commented May 12 at 17:00
  • It's this book that spawned this question. I've written to Gustafson but I don't think he checks his email. Commented May 16 at 12:11
4

You are possibly thinking of the 3M computer: 1MB of RAM, 1Mpx display, and 1MIPS.

2
  • 1
    No, its not that one. This had to do with the idea that you can't just have a fast computer, you need fast devices as well. Throughput vs. processing. Commented May 10 at 23:55
  • 4
    And it certainly predates data processing applications caring about 'display' performance.
    – dave
    Commented May 11 at 0:22
3

I seriously doubt Seymour Cray said (or believed) anything close to that.

The CDC 6600 was capable of about 3 megaflops, but had only 18 bit addressing, so it could only address 256K words of main memory.

The Cray 1 supported one megaword of RAM, but had a peak processing speed of about 160 megaflops.

The Cray 2 continued along the same general line--a maximum of 256 Megawords of memory, and peak processor speed of 1.9 gigaflops.

The Cray 4 supported up to 1 gigaword of memory and peak performance was 32 gigaflops.

Obviously I've skipped quite a few intermediate models (X-MP, Y-MP, etc.) but they did't change the basic balance much (if at all). None of them is anywhere close to a 1:1 ratio of flops to words of memory. The lowest ratio was the Cray 2, at about 8:1.

Although I'm not sure it was expressed in terms of the ratio of processing speed to memory, Seymour left Cray Research after the Cray 2 because he basically didn't like the direction the company was going. The rest of the company was trying to keep investors happy with relatively low-risk work like selling memory upgrades (which obviously reduce the flops:memory ratio). He wanted to do a radical new CPU design to get larger improvements in processor speed.

So, looking at it from the perspective of the ratio of processor speed to memory size, he considered 8:1 too low a ratio, and split off to work on the Cray 4 to push that ratio even higher (to 32:1 in the Cray 4).

If he'd looked at it from that perspective, I'd guess Seymour Cray would have advocated a minimum of a 10:1 ratio of flops to memory addressing (and seemed to prefer it even higher than that).

1
  • I recall the quote being in reference to a later machine, almost certainly the Cray-2. But looking at everything I have on that machine I don't see the quote. I have a copy of "The Supermen" somewhere, and I think that is where I saw it originally, but now I can't find that book. Commented May 16 at 12:09

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .