More news about HP's new memristor technology, from EE Times. Memristors are a new class of electronic device that change their resistance according to how much current goes through them - and, crucially, maintain that resistance when the current goes away. As the effect works on nanometric scales, this creates the possibility for very dense, very big, very simple memory chips -- a basic technology that gets any tech company salivating.
HP first announced memristors in April this year. It now claims that it's made a lot of progress very quickly, most notably in working out exactly how the things work. The basic idea is quite simple: titanium oxide changes its conductivity depending on how much oxygen is present inside its crystal structure, and vacancies within that structure - places where oxygen could be but isn't - can be moved around by electrical current. Now, HP says that the most interesting and important part of this effect occurs not in the body of the material as first thought, but at the interface between the titanium oxide and the electrodes that move current through it.
So far, so interesting. HP has found this out by learning how to build experimental devices in a way that lets it fine-tune lots of parameters and observe what happens, so it's building up a body of expertise in practical memristory. One early result is that the switching time for the devices is of the order of 50 nanoseconds - slower than dynamic or static memory, but much nicer than flash. This confirms HP's initial thoughts that memristor memory will be a natural contender for the flash market - and the company's saying that prototype chips may be available as soon as next year, which if true would be a world record lab-to-fab time for brand new physics. Normally, twenty years is closer to the mark. HP calls these chips resistive RAM or RRAM, although whether this is pronounced are-ram or with a gutteral roar like a small child imitating a racing car is not yet clear.
But wait - there's more. Because memristors are basically analogue devices, able to hold a wide range of values, they may be particularly apt for modeling analogue computation systems like synapses and self-learning systems such as neural networks. That's five years ahead for working systems, says HP, with another five years for commercialisation.
Got a good feeling about this one.