Wednesday, August 19, 2009

22 nanometer microtubes likely to extend Moore's Law indefinitely

IBM Taps DNA Structures to Extend Moore's Law

Posted by Arthur Cole Aug 17, 2009 8:00:15 PM

It's an interesting development, even if it probably won't have any practical applications for a dozen years or more, if that.

I'm talking about IBM's recent breakthrough in what some are calling "DNA-based chip technology." The company, along with researchers at Caltech, have figured out a means to arrange DNA-like structures on lithographic templates that are compatible with existing semiconductor manufacturing technology. The development is potentially significant because it offers the possibility of using carbon nanotubes in processes of less than 22 nanometers, the level at which most experts agree that current designs cannot cross.

If the design can be emulated in a practical fabrication environment, there is the possibility that we could see the continuation of Moore's law -- the doubling of CPU transistors every 18 months -- indefinitely.

The researchers describe their DNA structures as "origami," and say they can be used to create tiny scaffoldings on which millions of carbon nanotube could self-assemble into what are essentially tiny circuit boards. And the fact that this can be done on surfaces that won't require a wholesale revamping of the chip fab industry means it is likely to find some very willing backers.

Not everyone is jumping for joy, however. Assuming that the process proves to be commercial viable, there will still be quite a long period without any major advances in processing power, according to TweakTown's Sean Kalinich. Leakage on current 40 nm processes have already proven to be so severe that ATI and NVidia find themselves with serious yield problems. At this rate, current designs will have hit the wall long before the IBM scheme comes to fruition.

There's also a double-edged sword at play here, said tech blogger and self-described "crackpot and crank" Louis Savain. While this would dovetail nicely with the trend toward "brain-like neural networks" consisting of large numbers of parallel processors, it does nothing to address the memory bandwidth crunch that results from too many cores vying for too few data busses. If you think things are bad for memory systems coping with several hundred cores, try a few thousand or a few million. The only way to solve that problem is to figure out a way to embed numerous processors directly into the memory substrate.

Despite these concerns, however, the simple truth is that success is rarely total -- that every advancement will inevitably lead to new challenges. That there is now a chance that the 22 nm barrier can be crossed using existing fab technology is an achievement to be welcomed. And if there are additional challenges down the road, well that's what the design community is all about.

No comments:

Post a Comment