Re: 1teraflops cell processor possible?

On Thu, 27 Nov 2003 09:46:55 +0100, Jan C. Vorbrüggen
<[EMAIL PROTECTED]> wrote:

>> Not only am I convinced that the point the statement (which is a bit
>> of hyperbole) is trying to make essentially true, but I believe that
>> the reality it is summarizing is going to dominate the future of
>> high-peformance computation.
>
>Why "the future" - where, do you think, went the money in designing
>and building a Cray-2, say?
>
>> In a streaming architecture, you fetch data from somewhere, you do
>> something with it, you do something else with it, you do something
>> else with it, you do something else with it, ad nauseum, until you
>> have absolutely run out of operations you can stream together.
>> *Then*, and with great reluctance, you bear the cost of putting the
>> data where you will need to bear the cost of fetching it again.
>
>And I thought that was the whole point of optimizing compilers that do
>blocking and interprocedural analysis...?
>
Current microarchitectures have the following paradigm:

Get something. Do something to it. Put it back.

There is not a thing a compiler can do about that.  Between operations
(except in the case of an FMAC, which is an example of streaming),
operands have to sit in registers and they have to be ferried back and
forth.

The Merrimac authors were not referring to bandwidth to memory or even
to cache.  They were talking about bandwidth within the CPU itself.
In a streaming architecture there is no getting and putting; the
output of one functional unit feeds right into the input of another.

>> The people who make the budget decisions in
>> Washington are trying to get supercomputing back on track by paying
>> people to build big machines.  That's not an indefensible exercise.  I
>> just wish they would put more money into moving us away from a
>> computer architecture that is showing its age and its limitations.
>> Too much money is going into sheer size and not enough into finding
>> new ways of doing business.
>
>The problem will be that you will need both to be successful in this
>endeavour: you need a solution scaleable to sufficient size. Building
>small proof-of-principle demonstrators doesn't cut it.
>
Erg.  Japan has a world class vector processor, and the US is lashing
together thousands of Opterons and Power4's.

RM