Friday, August 11, 2017

IBM + Partners breathe new life into Moore’s Law with 5NM chip technology

By Rich Ptak and Bill Moran

When IBM exited the chip foundry business several years ago most industry watchers were sad to see the company go. A key player for decades in semiconductor research, IBM would definitely be missed. We thought that the industry had suffered a real loss. We and others thought statements of IBM’s commitment to making further investments in semiconductor research were to be written off as an essentially meaningless face saving gestures.

As it turns out, we couldn’t have been more wrong. In fact, IBM research continued the work on semiconductor research that it had been doing for nearly 50[1] years.  The IBM-organized consortium of IBM, Global Foundries, and Samsung based at NY State’s SUNY campus in Albany, is delivering a significant breakthrough in semiconductor technology research.  Exiting the chip foundry business was not a sufficient reason for IBM Research to cease its efforts.

Here’s some background on what IBM and its partners have accomplished. Moore’s law[2] says that the number of transistors on a chip will double approximately every two years. The results of that law drove the semiconductor industry for decades.
A meter being roughly a yard.  It might have been more fun if the industry has used the term “nanoyard”. However, as the topic is worldwide technology, the metric system is used, as is the practice in global technical and scientific circles.
Recently, much published commentary (ours and others) discussed how the law was reaching the end of its useful life. A major reason being the physical limits of chip geometry. Incidentally, one of the effects of law is that today’s cellphones (which fit in a pocket) have more processing power than the 1960’s computers used for the Moon visit (which occupied an entire very large room).

To understand what is going on, we need a little computer industry technology background. The industry initially measured processing speeds in seconds. Things moved faster so the term “milliseconds” (one-thousandth of a second) became standard. As the speed-up continued, the “microsecond” (one-millionth of a second) became the standard. One might imagine that things could not get much faster, wrong. Today’s process speeds are measured in “nanoseconds”, i.e. billionths of a second.   

Moving to semiconductors, chip size is measured in terms of the distance between identical features in an array.  The current unit for this distance is “nanometer”, i.e.  a billionth of a meter. The leading edge for productions semiconductor chips today is 10 nanometers.

Moore’s law depends on shrinking the size of the chips while increasing the number of transistors on the chip which increases processing power. Conventional wisdom was that it wouldn’t be possible to push the FinFET technology (which underlies chip manufacturing today) to much smaller chips without losing efficiency. Thus, the comments about the end of Moore’s Law.

However, IBM along with its partners have now developed a process around the Stacked Nanosheet Gate-All-Around Transistor.  This allows eventually building 5 nm chips with improved efficiency, which could not be achieved with FinFETs. The details of the process exceed the scope of this paper. (Those interested can start here[3].) The chart below provides a simplified view of the new IBM process compared with the existing industry standard 10 nm process.

Chart 1 This chart is an adaption of a copyrighted IBM chart.
There are several items of note. IBM's  new chips orient transistors horizontally versus today’s vertical arrangement. This allows transistors to be stacked and have more on each chip. Also, IBM chips use a new way to form the sheet material that is much more efficient in power consumption. It delivers a 75% saving in power compared to existing 10 nm chip architectures.  Finally, the new sheet formation process allows continuous fine-tuning for power and performance of specific circuits during manufacturing. Something not possible with FinFET technology

In summary, it appears to us that the IBM consortium has breathed new life into Moore’s law. With this new architecture, it looks to be applicable for the next decade or so just when many were pronouncing the law dead. However, we expect the interest and investments in such new technologies as quantum, data-centric computing and other approaches to grow.

[1] Publication in 1974 of a paper by Robert Dennard et al on MOSFET scaling rules for improving transistor density. See
[2] Not actually a law, it was a prediction about how the semiconductor industry would evolve in terms of density and cost per transistor.

No comments:

Post a Comment