Re: 1teraflops cell processor possible?

"Robert Myers" <[EMAIL PROTECTED]> wrote in message
news:[EMAIL PROTECTED]
> On Thu, 27 Nov 2003 13:56:51 +0100, Jan C. Vorbrüggen
> <[EMAIL PROTECTED]> wrote:
[SNIP]

> No, I don't think it is, either.  What bothers me is what I perceive
> to be an enormous misallocation of resources.  Too much money going
> into building big machines with yesterday's architecture, not enough
> going into more basic research on both hardware and software to find
> the best way to address problems in a massively parallel fashion.

On the other hand look at IA-64... Huge sums of money in
"tomorrows" architecture and yet those clunky old ones
seem to stay within touch of it on a tiny fraction of the
IA-64 budget.

The advantage of using stuff you already have to hand
when you're working on something new is that it's a known
quantity, it's almost certainly cheaper in the short-term
(important if you have limited funding), and it allows you
to prototype more, erm, accurately...

Not to mention the rigged-demo potential for the less
scrupulous out there...

> Yes, finding ways of programming something like Merrimac or a GPU to
> do real problems is going to take a massive amount of work and it will
> cost alot of money.  That money spent is an investment.  Once you gain
> the knowledge, you've got it forever.  Blue Gene will be scrap in a
> few years.
>
> And one thing seems certain: as feature sizes shrink, the energy cost
> of moving data around, even over very short distances, will dominate
> all other energy costs.  Even if energy cost nothing (and it
> doesn't!), getting the energy out of the CPU will be the limiting
> factor in how fast something like a protein-folding calculation can
> proceed.

Which is precisely where densely packed MPPs running low-
power cores can help, if your app fits. :)

I think that it will be very tough to beat MPPs over the
next few years if you value high throughput and low power
consumption. Yes moving data around off chip is bad, but
there are certainly are apps that can live with that
penalty.

Super-fast cores are just going to eat PSUs and air-con
for breakfast. Furthermore to do a given amount of work
they will *still* need to move data on and off chip anyway
(perhaps not as much as MPPs in many applications YMMV).
Ultimately it's up to the customer to decide if their app
is a good fit. I don't think there will ever be a one-size
fits all solution for this particular pissing-contest.

Don't forget that MPPs have been used to crack open new
problem spaces for many years now. They've come in all
shapes and sizes too - quite often to fit a particular
problem space. Therefore I would argue that you would be
misguided to assert some other approach as being more
original than a MPP design... :)

You should take a look at the MPP machines that have
been and gone and compare the variety in their designs
to vector machines, or boring old Minis, Mainframes
and Micros. I think you will be surprised. Of course
having lots of kinks doesn't mean something is
inherently good, just more interesting at least. :)

Cheers,
Rupert