June 30, 2011

Moore's law will not hit the wall at exaflop supercomputers

Some academics have written an 8 page paper about exascale computing and the end of Moore's law.

Marc Snir et al try to make the case that

* We will hit the end of Moore's law in mid-2020 with 7 nanometer lithography chips.
* spintronics may not be ready in time

they call for

* reducing the amount of communication used by computations.
* smarter use of CMOS circuitry
* inventory new device technology to determine which can be deployed in 10-15 years

I disagree on several aspects

1. I disagree that we will hit the wall at exaflop supercomputers. We can have an exaflop supercomputer by 2015. (NSA supercomputer).

2. Intel should have 7 nanometer process chips by 2017

IBM and have nanowires working at 2.6 to 3 nanometers.

Directed self assembly and nanoimprinting could get computer structures to 1-2 nanometers.

The academics make a case for the need for faster supercomputers. I think it is more important to get faster computers that use less power across the entire range of computers. Algorithms and methods should be developed to make the smallest computers more energy efficient to enable new forms of computers and sensors in spite of relatively slow progress with batteries and beamed power.

Having a battery that can last for 20 years because a device is made to operate cleverly to communicate over a distance of 45 kilometers across open wifi spectrum is an example of enabling technology for radically new applications.

A Wimax or other longer range version of such devices could enable reliable communication across entire countries (especially with some repeaters in skyscrapers or mountain tops or nanosats or blimps.)

Centrally and reliably managed superclouds of computer processing are interesting and a way to more fully utilize large amounts of processing power than dedicated supercomputers.

For supercomputers advancing science, I think there are hard optimization modeling for society and modeling to achieve and manage molecular nanotechnology. Also, having more computer power for better artificial general intelligence.

Quantum dot switches have been created that can perform femtojoule computing operations.

There is advancement on mass production of quantum dots and towards theory and experiments to develop computing around quantum dots.

There was Sub-femtojoule all-optical switching using a photonic-crystal nanocavity (journal Nature Photonics (May 2010)

There is progress towards femtojoule phase change memory (2009)

Femtojoule operations would mean one watt for a petaflop of processing and 1000 watts for an exaflop and a megawatt for a zettaflop. 100 zettaflop supercomputers would need 100 megawatts of power.

IBM is talking about getting beyond silicon with phase change memory and logic and nanophotonics (and nanoplasmonics after that).

Super lower power onchip photonics could enable zettaflop supercomputers with an architecture that is relatively similar to current practice.

Architectures will likely have to undergo more radical change beyond a zettaflop. There will be issues about communicating across any distance covered by the speed of light as faster operations are needed. More parallelism would reduce the need for faster operations.

Attoseconds and distances traveled.

Attoseconds are 10^-18 seconds. 1 attosecond is the time it takes for light to travel the length of three hydrogen atoms.
Zeptoseconds are 10^-21 seconds
Yoctoseconds are 10^-24 seconds. 1 ys: time taken for a quark to emit a gluon. The time that light needs to traverse an atomic nucleus

If you liked this article, please give it a quick review on ycombinator or StumbleUpon. Thanks
blog comments powered by Disqus