Moore's Law for Power Consumption: situation critical

THERE'S AN INTERESTING article at ACM Queue, called Making a case for Efficient Supercomputing by Los Alamos National Labs staffer Wu-Chun Feng.

It's time, says Wu-Chen, for the computing industry to start using alternative metrics to measure performance.

The reason? One is that "Moore's Law for Power Consumption", which he coins, is that power consumption of nodes doubles every 18 months. Meaning that when such nodes consume and dissipate extra power, not only must they be spaced out but they've got to be very aggressively cooled.

He shows a chart similar to one Intel's also shown in the past - if chips continue to heat up it won't be too long before we'll all have the equivalent of nuclear reactors inside our desktop boxes.

Los Alamos uses a 240-processor supercomputer that's about the size of a telephone kiosk but takes less than 5.2KW of power at full load without aircon, without ventilation, and without air filtration.

A high end 240 processor supercomputer like a Beowulf cluster needs a specially cooled machine room, and eats over 36KW of power.

Green Destiny uses Transmeta server blades, with each CPU getting rid of 75 per cent of the trannies used in a traditional RISC architecture.

While he says that Transmeta CPUs are "significantly more reliable" than conventional mobile processors, its floating point performance is its Achilles' Heel. So Los Alamos changes its code morphing software and made a better performance CMS that improves the FP by nearly 50%.

There's a lot more here. µ