UC Berkeley Professor Leon Chua deduced by mathematical symmetry arguments that there was a missing basic passive electronic circuit element back in 1971. The basic passive circuit elements create physical relationships between the four electromagnetic quantities:  charge, current, voltage, and magnetic flux. Besides the resistor, capacitor, and inductor reasoned Chua, there should be another element that connects charge and flux. He named the missing element the memristor, which becomes more or less resistive based on the amount of current passed through the element. The resistance change is long term; the memristor has memory; hence the name. However, HP's device does not use magnetic flux as Chua's theoretical memristor did. It doesn't store charge like a capacitor. Instead, HP's device alters its resistance dependening on the history of the currents passing through it using an atomic-level restructuring within it's 5nm resistive titanium-dioxide layer. The chemical restructuring results in movement or redistribution of oxygen vacancies within the titanium dioxide which in turn produces a resistance change in the material.

Chua’s was just a mathematical argument based on the existing differential equations for the three common passive circuit elements. It took nearly another four decades for HP Labs researchers to build a practical, working memristor using nanoscale imprint lithography, 50nm platinum wires, an insulating layer of titanium dioxide between the wires, and semiconductor processing. The HP researchers published in Nature magazine but you can read about it online in this article in Scientific American and in this article on Arstechnica, which says that HP researchers have already achieved storage densities of 100 Gbits/cm², although I think the latter is misrepresenting the achievement as though HP had actually built 100 billion memory bits on one chip when the image actually shows 17 devices. However, the achieved memory density is already 5x the current storage density of Flash memory chips, and that’s for an experimental circuit. Impressive. (Queue Darth Vader-like voice.)

One blog entry I read online at CNET quotes Stan Williams, who heads up the Information and Quantum Systems lab at HP Labs, as saying that the fabricated memristor’s switching speed was “faster than we can measure,” although the New York Times seems to report just the opposite. Too fast to measure—that’s always a good characteristic to build on when you’re dealing with a memory element. So you get density and speed with this nascent nonvolatile memory technology. It’s starting to sound like unobtanium. The Wikipedia entry for the memristor claims that the name “flux capacitor” could be an accurate name for the memristor. Shades of the movie Back to the Future, although I doubt that you’ll be able to combine memristors, DeLorean automobiles, and plutonium to build a time machine.

Because HP’s memristor has long-term memory and is built with nanoscale imprint lithography, you can easily envision using the technology to make large amounts of nonvolatile memory on future semiconductor memory chips, ASSPs, and SOCs. Because it’s an analog device, you can envision doing other things with the memristor as well. One such suggested application was real, large-scale, analog neural networks (like the ones in our brains, as opposed to the digital simulations we presently build).

However, the odds suggest it’s more likely to take a while before memristors go big-time commercial, if they ever do. Far-fetched nonvolatile-memory schemes for semiconductors have a sketchy history: magnetorestrictive RAM (MRAM), ferroelectric RAM (FRAM), and magnetic bubble memory to name a few. (Yeah, I know that bubble memory’s synthetic garnet isn't a semiconductor, but the stuff was way cool and even Intel got into that game—and then out when it petered out commercially.) I have written about many of these experimental memory technologies in EDN over the past 20+ years and most failed to produce competitive memory products. The same was true for EEPROM, as I recall, until it morphed into block-erasable Flash (it erases in a flash, get it?). Nobody’s laughing at Flash memory any more. In fact, our digital civilization is heavily dependent on the Flash memory in our mobile phones, digital cameras, music and personal-media players, USB memory sticks for laptops, and in many, many embedded devices.

It may turn out that the memristor’s titanium dioxide or platinum are not compatible with production semiconductor manufacturing, although the quick reads I've done suggest otherwise. Certainly we've recently seen new materials being quickly taken up in the semi process flow including copper for interconnect (copper's too highly reactive to work in ICs, I recall reading) and now the transition metal hafnium for a hi-k dielectric to ameliorate the static leakage problems with silicon dioxide at 45nm and below. It may also turn out that there are better materials to use for memristor fabrication than platinum and silicon dioxide. No one really knows just yet.

HP's announcement admittedly describes a science project. There are no memristor memories in the immediate future. As to why HP published this information at this stage, several reasons come to mind. HP Labs has a very long history of publishing such developments. Only HP has rigorously published its scientific and engineering developments in its own HP Journal, which is now dead and gone (but fondly remembered). But the publishing gene remains strong at HP Labs. I'm sure HP's filed the necessary patents before the announcement in Nature. Now they need to find a manufacturing partner, since HP divested itself of all its production fabs in the Agilent spin off (later spun off again into Avago and subsequently purchased by Marvell in 2006). Certainly if I owned a fab or foundry, I’d be investigating HP’s science experiment to see what it might become.