K said:
Robert Myers wrote:
Exactly. Off-the-shelf is "cheap". ...even if it doesn't work. ;-)
Is it too optimistic to imagine that we may be coming to some kind of
closure? That you can do so much with off-the-shelf hardware is both an
opportunity and a trap. The opportunity is that you can do more for
less. The trap is that you may not be able to do enough or nearly as
much as you might do if you were a bit more adventurous.
It apparently didn't take too many poundings from clusters of boxes at
supercomputer shows to drive both the customers and the manufacturers of
big iron into full retreat. The benchmark that has been used to create
and celebrate those artificial victories was almost _designed_ to create
such an outcome, and the Washington wizards, understandably tired of
being made fools of, have run up the white flag--with the exception of
the Cray X-1, which didn't get built without significant pressure.
I'm hoping that AMD makes commodity eight-way Opteron work and that it
is popular enough to drive significant market competition. Then my
battle cry will be: don't waste limited research resources trying to be
a clever computer builder--what can you do with whatever you want to
purchase or build that you can't do with an eight-way Opteron?
The possibilities for grand leaps just don't come from plugging
commodity boxes together, or even from plugging boards of commodity
processors together. If you can't make a grand leap, it really isn't
worth the bother (that's the statement that makes enemies for me--people
may not know how to do much else, but they sure do know how to run cable).
Just a few years ago, I thought commodity clusters were a great idea.
The more I look at the problem, the more I believe that off the shelf
should be really off the shelf, not do-it-yourself. It's not that the
do it yourself clusters can't do more for cheap--they can--they just
don't do enough more to make it really worth the bother.
Processors with *Teraflop* capabilities are a reality, and not just in
artificially inflated numbers for game consoles. Not only do those
teraflop chips wipe the floor with x86 and Itanium for the problems you
really need a breakthrough for, they don't need warehouses full of
routers, switches, and cable to get those levels of performance.
Clusters of very low-power chips, a la Blue Gene was not a dumb idea, it
just isn't bold enough--you still need those warehouses, a separate
power plant to provide power and cooling, and _somebody_ is paying for
the real estate, even if it doesn't show up in the price of the machine.
_Maybe_ some combination of Moore's law, network on a chip, and a
breakthrough in board level interconnect could salvage the future of
conventional microprocessors for "supercomputing," but right now, the
future sure looks like streaming processors to me, and not just because
they remind me of the Cray 1.
Streaming processors a slam dunk? Apparently not. They're hard to
program and inflexible. IBM is the builder of choice for them at the
moment. Somebody else, though, will have to come up with the money.
RM