Newsgroups: comp.parallel
From: alberto@moreira.MV.COM (Alberto C Moreira)
Subject: Re: Massively Parallel "Pizza Box"
Organization: MV Communications, Inc.
Date: 2 Oct 1995 14:47:19 GMT
Message-ID: <44ou1n$2jo@usenet.srv.cis.pitt.edu>

In article <44hai9$62l@usenet.srv.cis.pitt.edu> eugene@pioneer.arc.nasa.gov (Eugene N. Miya) writes:

[some deleted...]

>Paraphrasing Ivan Sutherland:
>        a CPU can only really work on one word of memory at a time.
>        placing a bigger memory on a CPU means N-1 other words aren't
>        being used when N is getting larger.  There must be a better
>        way.

       If I'm fast, I don't care how many bits are left unused. A 4Mb DRAM 
       costs a couple of bucks or so; why worry ? The same applies to
       processors: when a PE costs $2, I won't care about efficiency. In
       my view, N can get as large as it possibly can, and as long as the
       cost is well bounded by some affordability function, I don't think
       optimizing memory usage is very relevant.

>I leave out long standing ideas like PIM because these are merely
>different scalings of the same problem.  Hillis was one of Ivan's
>students.

       If I have enough processors, even the Von Neumann bottleneck only
       applies to one of them at a time. The problem becomes not so much
       how fast a processor runs, but how good the overall architecture is.
       And yet, the fastest the individual PE the better. 

       I don't think "scaling" is just a "merely" sort of thing. In real life,
       big oh notation just isn't tight enough. A processor that's twice as
       fast may make a 48-hour job doable in 24-hours, for example,
       weather forecasting. I'll take any scaling I can get, as long as my
       computers get faster - and no more expensive.

       If one would do Hillis' machine today using 1995 chip technology, the
       whole array of PEs would probably fit in a handful of chips, for a
       fraction of the cost, and probably at a much faster clock rate.

[some more deleted...]

>>            Right now there's still plenty of mileage to be achieved
>>            by scaling down chip dimensions. Even within the current
>>            architectural limits, there's no reason why we can't increase
>>            our clock rates by a fair amount,

>Quantify please?  Please justify your answer.  Picoseconds, femtoseconds?
>Explain the limitation.

       The SHARC is clocked at 40 Mhz. A cheapie Pentium is clocked at 133 Mhz,
       an Alpha over 200Mhz. Geometries have been falling from .8 down to .6 
       and .5 micron; I just read an ad, I don't remember who, claiming that 
       their next generation chips will be etched at .25 micron.

       You do the arithmetic and see that the limitation, at least 
       right now, is time for the technologies to mature. There's more 
       electronic technology out there than able-minded professionals 
       to develop it into products.

>I just finished reading a good book:
>        Filters against Folly by biologist G. Hardin
>My comment to you on this is (from the book)
>        There is no such thing as a free lunch.

        True, but people will push against the boundary as much as they can.
        So far we haven't been disappointed, and I still see some mileage 
        ahead. Why be pessimistic ?

>SO what if the readers of this group chipped together $99, you will build
>us a 64-node parallel machine of unknown topology with unknown software (NT?)?

        If somebody gives me a sound financial backing, yes, I'll take the 
        challenge. For $99 you won't get more than a bunch of 8051s on a
        board; but scale the money proportionately to the cost of good 
        chips - the SHARC is one - and you'll see that it's doable. And the
        moment parallel technology hits the mainstream PC market, volume
        development will start and prices will plummet.

>I can afford $1.  I just donated $1,000 to a local university.
>So you build it, then what?  It's going to run my users' CFD codes?

        If I build it, it'll target the commercial market. Spreadsheets, 
        Graphics, Sound, Video; that's where the money is. A PCI
        board or two full of PEs that can plug into a PC. Small is
        beautiful...

[some more deleted...]

>>            When that happens, today's concept of "supercomputer" will
>>            be as obsolete as the concept of mainframe is today.

>Oh?

        Just wait. Just wait... Even today, it already sounds like COWs are
        just as good as many an expensive supercomputer. Technology
        doesn't stop!


                                                           _alberto_

