|
shallowdeep -> RE: IBM builds Carbon Computer. (11/3/2012 12:52:07 AM)
|
quote:
ORIGINAL: Moonhead
True objections, but as a first attempt, it bodes rather well, doesn't it?
Yes, it does seem promising; however, it's one of those technologies where a practical application, if it comes, is probably still a ways away – perhaps ten years or so. Around 90% of the targeted placement sites had CNTs adhere to them and, of those, around 90% were functioning semiconducting devices (although it seems only around 58% performed as expected). As an early step this is neat, but there is still quite a hurdle to get from the current yield to rates that would work when integrating transistors into interdependent logic circuits useful for computation. With conventional processor designs, even a single failed transistor can ruin a logic core that consists of tens of millions of the things.
In the paper, most of the failures of connected CNT devices were attributed to the presence of undesirable, conducting metallic nanotubes. Improvements in separation techniques can probably significantly reduce those failures. Getting the tubes to bind to a higher percentage of the target sites might be more difficult, especially as the process is scaled down from the older, relatively large one used for the research toward sizes that would be more competitive with commercial processes (currently logic is at the 22 nm node, and structurally simpler flash memory is as low as 19 nm). However, fundamentally, the approach is probably never going to have device yields on par with conventional CMOS. That might not be the end of the world: there are ideas for redundant defect-tolerant architectures and other approaches to minimize the impact of defects on chip yields, but they require significant rethinking of designs and come at a cost in wasted die and power. Given that the CNT-based FETs might offer something around a 5x improvement in performance over silicon, that begins to chip away at the advantage.
A nice thing about the technique is it largely leverages current lithography fabrication approaches, so it could be easier to get into production if some of the significant issues can be worked out. If it continues to show promise, something similar just might be used to make processing chips by the end of the decade.
|
|
|
|
