Re: 1teraflops cell processor possible?

On 01 Dec 2003 09:03:19 -0800, David Gay
<[EMAIL PROTECTED]> wrote:

>
>Robert Myers <[EMAIL PROTECTED]> writes:
>> If you believe that joules per flop is the ultimate limiting factor,
>> then, yes, movement of data on the chip is the performance limiter.
>> Movement from a functional unit to a register accomplishes nothing,
>> and, as feature sizes shrink, the energy cost of moving the data will
>> exceed the cost of performing the arithmetic.
>
>The argument that computation power is limited by flops seems bogus to
>me. It would lead to the conclusion that all the work in, say, sorting is
>in the comparisons, and that the permutation is inherently free. Or, to
>take another example, compilers spend most of their time creating data
>and moving it around. Does that mean that compilation has potentially
>zero energy cost?
>
Oddly enough, even though flops are the wrong unit to be using when
talking about calculations that are not floating-point intensive, the
argument goes though the same way.  Just change the sentence to:

If you believe that joules per useful unit of work done is the
limiting factor, then movement of data on the chip is the performance
limiter.

And, no, the conclusion is not that compilation (or other similar
tree-traversing activities, like AI) will be free.  Quite to the
contrary, it says that those kinds of applications will be more
resistant to low-energy strategies than traditionally power-hungry
floating-point intensive calculations, where, more often than not, you
know how to exploit code and data locality much more effectively.

RM