The J Curve

Sunday, September 26, 2004

Transcending Moore’s Law with Molecular Electronics

The future of Moore’s Law is not CMOS transistors on silicon. Within 25 years, they will be as obsolete as the vacuum tube.

While this will be a massive disruption to the semiconductor industry, a larger set of industries depends on continued exponential cost declines in computational power and storage density. Moore’s Law drives electronics, communications and computers and has become a primary driver in drug discovery and bioinformatics, medical imaging and diagnostics. Over time, the lab sciences become information sciences, and then the speed of iterative simulations accelerates the pace of progress.

There are several reasons why molecular electronics is the next paradigm for Moore’s Law:

• Size: Molecular electronics has the potential to dramatically extend the miniaturization that has driven the density and speed advantages of the integrated circuit (IC) phase of Moore’s Law. For a memorable sense of the massive difference in scale, consider a single drop of water. There are more molecules in a single drop of water than all transistors ever built. Think of the transistors in every memory chip and every processor ever built, worldwide. Sure, water molecules are small, but an important part of the comparison depends on the 3D volume of a drop. Every IC, in contrast, is a thin veneer of computation on a thick and inert substrate.

• Power: One of the reasons that transistors are not stacked into 3D volumes today is that the silicon would melt. Power per calculation will dominate clock speed as the metric of merit for the future of computation. The inefficiency of the modern transistor is staggering. The human brain is ~100 million times more power efficient than our modern microprocessors. Sure the brain is slow (under a kHz) but it is massively parallel (with 100 trillion synapses between 60 billion neurons), and interconnected in a 3D volume. Stan Williams, the director of HP’s quantum science research labs, concludes: “it should be physically possible to do the work of all the computers on Earth today using a single watt of power.”

• Manufacturing Cost: Many of the molecular electronics designs use simple spin coating or molecular self-assembly of organic compounds. The process complexity is embodied in the inexpensive synthesized molecular structures, and so they can literally be splashed on to a prepared silicon wafer. The complexity is not in the deposition or the manufacturing process or the systems engineering.

Biology does not tend to assemble complexity at 1000 degrees in a high vacuum. It tends to be room temperature or body temperature. In a manufacturing domain, this opens the possibility of cheap plastic substrates instead of expensive silicon ingots.

• Elegance: In addition to these advantages, some of the molecular electronics approaches offer elegant solutions to non-volatile and inherently digital storage. We go through unnatural acts with CMOS silicon to get an inherently analog and leaky medium to approximate a digital and non-volatile abstraction that we depend on for our design methodology. Many of the molecular electronic approaches are inherently digital and immune to soft errors, and some are inherently non-volatile.

For more details, I recently wrote a 20 page article expanding on these ideas and nanotech in general (PDF download). And if anyone is interested in the references and calculations for the water drop and brain power comparisons, I can provide the details in the Comments.

7 Comments:

  • Steve,

    This was finally the post on your blog (which I usually read to get my share of out-of-this-world ideas ;-) ) about which I do actually have a rather educated opinion. I'd love to read your rebuttal! ;-)

    I guess that everyone here agrees that we need to find some other than CMOS technology to fuel our expanding computing/telecommunication needs. I'd move the Power problem up a notch (cooling requirements => chip packing density => system size L => communication latency of L/c, the 'c' being, unfortunately, fixed). Also, one does not see true 3D integration in semiconductors because those little transistor critters like to grow on pretty well-defined monocrystral surface, something that is really hard to grow over a couple levels of wiring, however planarized. And yes, then the thing will just melt...

    My main problem is with your statement that There are several reasons why molecular electronics is the next paradigm for Moore’s Law, the next part, not eventual. As you are writing in your paper, "Of course, we use these computers to make better design software and manufacturing control algorithms. And so the progress continues.". Are you sure that today's (or tomorrow's) computers actually have the power to "compile" my high-level description of my math problem into a protein molecule which will be able to tell me the answer? No, "... no computer has ever been built that is powerful enough to solve them. Even today's most powerful supercomputers choke on systems bigger than a single water molecule." Of course having access to a QC is an interesting option, but as far as I understand one almost needs a QC to compile a (non-trivial, i.e., not Shor's) algorithm into that.

    Which brings me to my point, supported by SIA's roadmap's Emerging Research Devices section -- Superconductor Electronics/SFQ/Josephson junction devices. Yes, I am biased ;-), has been working in this particular field for maybe 15 years, up until recently. Couple years old SIA roadmap listed RSFQ as "least risk" next-generation technology (the current one does not have this field). Yes, it is "quantum" as in dissipated power achieves quantum limits and the logic is based on naturally quantized (though still digital) behavior. Yes, the best part is that "it is not nano-, it is micro-", but maybe it is also the worst part for obtaining federal funding for its development. Some of us are still trying to push this technology forward, to bridge the gap between current CMOS (or III-V) devices and future molecular/QC ones, we will see how this plays out.

    Paul B.

    (pbunyk@lycos.com)

    P.S. D-Wave is building a quantum computer using aluminum-based circuits. -- sorry, those are is niobium-based circuits! ;-) A tiny bit of Al is left there as a byproduct of obtaining aluminum oxide for Josephson junction tunnel barrier. But I do wish those D-Wave guys the best luck, even have some personal reasons to see them being around! ;-)

    By Blogger Paul B., at 7:12 PM  

  • Paul: Thanks. A few quick thoughts:

    1) D-Wave’s response to your P.S.: “Our qubits are made of aluminum. Only the control/readout circuits are made of niobium.” Full disclosure: I also have “reasons to see them being around” as I am on the Board. =)

    2) RSFQ is interesting, and there are several groups in Europe working on it, but I don’t think it’s the “next phase of Moore’s Law” in any mainstream sense. As we look at the list of needs in my original post, you seem to agree that “size” is a problem; interconnected SQUID loops, each with two or more Josephson junctions will not be pushing the envelope on size. The main problem is manufacturing cost and complexity. RSFQ involves esoteric manufacturing that is deeply embedded within the current process flow. Look how long it took something as “simple” as copper wiring to be adopted by the semi industry (given the interdependence with other process steps and materials). Reliability and capital cost concerns will retard the pace of progress, as it has for the last 30 years (e.g., the U.S. petaflop project).

    Molecular memories, in contrast, have a much more pragmatic path to near-term revenue. They are a “late binding” addition, like a low-temp passivation layer, that does not need to introduce change to the earlier process steps. These products are also well down the path to commercialization. Megabit test chips have been cycled to a trillion read/write cycles. Manufacturing partners are working to prepare for volume production.

    3) QC: I am not sure if I understand your concern about quantum computers. There have been a number of “software” breakthroughs in the past five years, from Shor to Grover to molecular modeling with adiabatic approaches, and most recently, for the solving of any partial differential equation. There appears to be a Moore’s Law-like doubling in the number of solid state entangled qubits over time. It is early still, like when Moore made his first observation in 1965. In the future, quantum computers will have a major impact on materials simulation and high end computing, but they are not a near term substitute for the dirt-cheap proliferation of computation that is driving so much of the economy today. You might like this discussion of quantum computation.

    By Blogger Steve Jurvetson, at 10:42 AM  

  • Steve,

    1) D-Wave’s response to your P.S.: “Our qubits are made of aluminum. Only the control/readout circuits are made of niobium.”I stay corrected. From one of
    their papers
    : "...an Al flux qubit coupled to a Nb resonant tank. The latter played dual, control and readout, roles." I was thinking about different qubits, made of "regular" Nb-trilayer junctions.

    Since you caught me on this, I got an half-an-hour lecture over the phone about relative merits of Al vs. Nb qubits from of the people affiliated with them, who really knows this stuff, which, I guess, was enough of a penance for my ignorance. ;-)

    Full disclosure: I also have “reasons to see them being around” as I am on the Board. =) Sure I realize this pretty well! ;-) The fact being one of the reasons why I value your opinion on those matters. Otherwise I'd be reading /. ;-)

    RSFQ is interesting, and there are several groups in Europe working on it

    Sigh... Take it from someone on the inside of this particular crowd. European SFQ groups are all in Universities and they were making great science, but no one there is serious about treating this as technology (with one exception that I know, Chalmers Uni., but that particular group is pretty much aligned with the US efforts). European guys are much more interested in getting their Ph.D.s studying some exotic single junction physics rather than treating SFQ from electrical engineering perspective and trying to actually build something.

    Japan is different, they do care about having access to (almost) real production fab. Too bad that all the big concerns there decided to merge their SCE groups under an umbrella of a single government organization.

    And the US -- well, there were two almost-commercial outfits working on trying to build real things up until recently, a small company and an R&D group in a large company, the latter was (in my obviously biased ;-) opinion) much more advanced and better managed, too bad that powers to be decided to pull the plug on us.

    interconnected SQUID loops, each with two or more Josephson junctions will not be pushing the envelope on sizeThere is absolutely no fundamental reason why JJs can not be as dence as CMOS transistors, if fab is done not on 10 year old 1 um resolution machines but on anything capable of at least 0.35 um (Nb circuitry does not really have to go further down in size to achive max. performace).

    RSFQ involves esoteric manufacturing that is deeply embedded within the current process flow.No, manufacturing is much simpler than anything else, basically thin-film technology. Also, it is so simple (and dissipated power is so small) that it is possible to seriously talk about real on-chip 3D integration, something that semiconductors can not do (transistors ARE formed on the wafer surface).

    Look how long it took something as “simple” as copper wiring to be adopted by the semi industryYes, because Cu offered resistance only a factor of 1.6 lower than Al, while comparable speed increases could be realised by architectural tricks people did not care much about copper wiring. Sufficiently cold Nb offers a factor of infinity lower resistance... ;-)

    Reliability and capital cost concerns will retard the pace of progress, as it has for the last 30 years (e.g., the U.S. petaflop project).Reliability -- maybe, because it was not really systematically studied for josephson-based devices. One thing that is known is that they are pretty damn rad-hard. ;-) Capital cost -- well, almost a dream superconductor foundry can be put together for about $30M, at least a factor of 100 less than modern $3B semi fab.

    As to the US petaflops (HTMT) project -- take it from someone who put together (manually! as an ex-HP R&D chip engineer you might get a chuckle ;-)) a 64K JJ 20GHz "microprocessor" chip demo on the very last drop of HTMT $$ and had no more left to properly test it, not to mention re-spin -- that project's problems were not exactly all technical.

    Finally,

    I don’t think it’s the “next phase of Moore’s Law” in any mainstream sense.I guess that we can say that QC is not "mainstream" in pretty much the same way, "they are not a near term substitute for the dirt-cheap proliferation of computation."But someone, somewhere would always want to build top-of-the-line system pushing the envelope of what is doable, and in my mind SCE has a pretty good chance to succeed in that, even if for smallish niches.

    Sincerely,

    Paul B.

    By Blogger Paul B., at 2:10 PM  

  • > The future of Moore’s Law is not CMOS
    > transistors on silicon. Within 25 years, they
    > will be as obsolete as the vacuum tube.

    You're clearly not an electric guitarist!

    By Anonymous Anonymous, at 11:04 PM  

  • or a tube-amp audiophile.... =)

    That's why I said "as obsolete." There will still be some applications for CMOS silicon transistors (and vacuum tubes), but they won't be the mainstream or the majority.

    By Blogger Steve Jurvetson, at 8:48 AM  

  • This is very interesting. I can only imagine how this will change our society. Hopefully, this will be better for the planet (only 1 watt to power all the Earth's computers! wow!) and will drop down prices, allowing third world countries to adquire more technology.

    Industrial desing would change a lot too.

    The water drop comparison is impressive. Thanks for the link, Steve.

    By Blogger Nell, at 1:08 AM  

  • Nell: thanks. Here are some further thoughts on the societal impacts of these advances.

    And for those who missed it: In a wonderfully symbolic nod to the future, Moore recently made a sizable donation to Caltech to “establish the Nanoscale Systems Initiative (NSI). The grant will support one of the scientific and technological community's promising research avenues--the creation of extremely tiny devices to augment and in some cases displace the state-of-the-art electronic systems of today.” (press release)

    All this, and he’s a great Salmon fisherman too. Here’s a photo of the catch.

    By Blogger Steve Jurvetson, at 11:21 AM  

Post a Comment

<< Home