the REAL Project Ultraviolet thread - SGI

This is a discussion on the REAL Project Ultraviolet thread - SGI ; So, did anyone what to have a non-pc/non-irix related discussion. one that actually was about ultraviolet? ..... is perhaps John Mashey still about? http://www.sgi.com/features/2004/jul...t_ultraviolet/ just to kick it off; MHO Dafydd Williams wrote: > On 2004-08-05 22:29:13 +1000, Alexis Cousein ...

+ Reply to Thread
Results 1 to 3 of 3

Thread: the REAL Project Ultraviolet thread

  1. the REAL Project Ultraviolet thread

    So, did anyone what to have a non-pc/non-irix related discussion. one
    that actually was about ultraviolet?
    ..... is perhaps John Mashey still about?

    http://www.sgi.com/features/2004/jul...t_ultraviolet/

    just to kick it off; MHO

    Dafydd Williams wrote:

    > On 2004-08-05 22:29:13 +1000, Alexis Cousein

    said:
    >
    > > VAN and Scalable Graphics are alive and well, thank you.

    >
    > Alexis, any chance of you shedding ~any~ hope on this little
    > discussion? SkyWriter, bless his heart, is depressing Ian. We can't
    > have a depressed Ian, now, can we? Where would us SGI hobbyists and


    awe heck, it's not that bad:
    1) Intel is popular! so use it (but not gcc - shudder)
    2) vector processing is always nice to have (well if you need it, and
    it's
    big enough to do any real work)
    3) cpu-in-memory is not a bad idea (even I have a couple of patents on
    it
    4) reconfigurable computing? sorry, i have always problems with this:
    a) even a hardware designer doing hand layout with pragma coding has
    a
    hard time wringing performance
    out of an FPGA; thinking a software guy can produce good results
    it
    is wildly optimistic.
    b) FPGA's change every year, will your code be portable?
    c) sounds more like a library of preconfigured special blocks picked

    ala cart on a general purpose bus
    internal bus is what wil be offered. this is usually a poor
    performer. maybe they license tensilica
    like technology with extensible you-do-it instruction sets.

    i don't think it's all that visionary. and I wouldn't credit SGI with
    the
    innovation, it's all been done before.
    i would LIKE to see it work; because i always thought it would be good
    (assuming the REAL technology
    behind it was there: reallyreallyreally smart compilers, and
    run-time-heretogeneous-resource allocation),
    but i'm not convinced this is the year.... much less the decade.

    >
    > fanboys be without futuretech?
    >
    > best always,
    >
    > Dave @ ATMOS
    >
    > --
    > Dafydd Williams
    > ATMOS Software Productions Pty Ltd
    > pipeline-at-atmos-dot-com-dot-au







  2. Re: the REAL Project Ultraviolet thread

    SkyWriter writes:

    > So, did anyone what to have a non-pc/non-irix related discussion. one
    > that actually was about ultraviolet?
    > .... is perhaps John Mashey still about?


    Given your penchant for name calling I'm pleasantly surprised that
    there's actually some higher cognitive functions at work too.

    I agree with most of your points, reconfigurable FPGAs in particular
    have been touted many times by bogus outfits like Starbridge and
    others, but as with IPF they all want to shift the hard work to
    software without first showing that it's even feasible in practice.

    However, mixing vector and scalar CPUs on a chip is all well and good
    but what it comes down to is, as always, memory bandwidth and latency.
    If you have the bandwidth to feed a decent vector unit the system is
    IMO very likely to be so expensive that, even though the scalar CPU
    will benefit a lot, those with scalar needs will get much better
    price/performance elsewhere. As always there are exceptions of course.

    PIMs, sure but will they be cheap and general enough? If you make a
    "PIM-cache" it's just a slight improvement, with memory prices
    starting to dominate system cost (perhaps with the exception of the
    really high end), special memory in volume is probably going to cost
    an arm and a leg. But here's hoping...

    *p


  3. Re: the REAL Project Ultraviolet thread

    Per Ekman wrote:

    > SkyWriter writes:
    >
    > > So, did anyone what to have a non-pc/non-irix related discussion. one
    > > that actually was about ultraviolet?
    > > .... is perhaps John Mashey still about?

    >
    > Given your penchant for name calling I'm pleasantly surprised that
    > there's actually some higher cognitive functions at work too.


    take linux and peecee's off the table, and i'll devote some time to it,
    rather than spend it elsewhere.

    >
    > I agree with most of your points, reconfigurable FPGAs in particular
    > have been touted many times by bogus outfits like Starbridge and
    > others, but as with IPF they all want to shift the hard work to
    > software without first showing that it's even feasible in practice.


    as design complexity grows the ability to get a product 'correct' in time
    for profitable production is getting harder and harder. there isn't enough

    time to either get the spec right, or verify the design (this is primarily
    hardware
    i'm speaking of). I assume the intent of the FPGA portion is to extend the

    instruction set of the processor 'complex' in one or two directions where

    it makes sense. such as; DSP function, or similar process not already
    provided by vector unit, ad hoc optimization of repeatable inline code,
    extension of PIM functionality.

    i do not see the ability to utilize this is any great user customizable
    way.

    >
    >
    > However, mixing vector and scalar CPUs on a chip is all well and good
    > but what it comes down to is, as always, memory bandwidth and latency.
    > If you have the bandwidth to feed a decent vector unit the system is
    > IMO very likely to be so expensive that, even though the scalar CPU
    > will benefit a lot, those with scalar needs will get much better
    > price/performance elsewhere. As always there are exceptions of course.


    scalar, and vector is almost always good (for the right cost), except for
    quake (i think)

    >
    > PIMs, sure but will they be cheap and general enough? If you make a
    > "PIM-cache" it's just a slight improvement, with memory prices
    > starting to dominate system cost (perhaps with the exception of the
    > really high end), special memory in volume is probably going to cost
    > an arm and a leg. But here's hoping...
    >


    it wasn't to clear about what the implementation of the PIM was. however,
    i assume it's processor in main memory, and not the processor in the
    memory chip
    .....err... scam.
    DRAM memory latency has not decreased by more than a factor of 3 in a
    couple of
    decades while the bandwidths attainable for the same 'data
    width/incremental
    controller cost' has increased by roughly 30. bringing a computational
    element
    close to memory (even a small one) reduces effect of that latency, as well
    as frees
    up bandwidth for the rest of the system.

    >
    > *p



+ Reply to Thread