Today, HP announced a notable development in its efforts to design memristor technology. The company has managed to map the chemical changes and basic structures of its own prototypes. The startling implication is accurate—HP managed to build a memristor before it fully understood how its own design functioned. The memristor was theorized to exist by one Leon Chua in 1971, but remained a hypothetical fourth circuit element (capacitors, resistors, and inductors are the other three) for decades. HP Labs finally managed to build a functional memristor in 2008.
Memristors are a 'Holy Grail' technology—they could change computing nearly as much as the invention of the transistor. The word 'memristor' is a portmaneau of "memory resistor." Unlike DRAM, memristors are non-volatile—they do not require constant electrical power in order to refresh data. This vastly simplifies the design of any theoretical memristor memory device. Since memristors don't require constant refreshes, cell densities can be much higher. The non-volatile nature of the storage combined with fast switching theoretically allows memristors to replace both DRAM and
primary storage at the same time.
An array of 17 oxygen-depleted titanium dioxide memristors. Each is about 50nm wide.
Consider the ramifications of that. Up to now, computers and the software they run have been designed around a tiered storage concept. When a CPU needs data, it checks L1 cache first, immediately followed by L2 and L3 (if applicable). RAM is next, with main system storage last of all. This remains true, despite the fact that SSDs are much faster than hard drives—there's a huge gap still between RAM
. Memristors—HP has dubbed the new design ReRAM— could all-but-eliminate that gap. To date, HP had difficulty analyzing the switching process because the channel where the switch takes place is tiny—only about 100 nanometers wide.
Highly focused X-ray bursts have proven capable of capturing the process. According to Stan Williams, a senior fellow at HP, the new discovery will significantly improve the quality of eventual products. We were on a path where we would have had something that works reasonably well, but this improves our confidence," Williams said. "[It] should allow us to improve the devices such that they are significantly better."
Because memristors remember their own charge states, they can be used to create self-programming devices. Amplifying the charge of Memristor A can begin a series of changes that program an entire array of memristors to perform Task B. Change the charge of Memristor A, and we change the programmed task. This has implications for both neural networks and simulated AI, and it may allow computer scientists to improve / simplify computer logic.
The first memristor-based equipment that comes to market will still use a significant amount of 'traditional' silicon—even once we've managed to commercialize memristor products, the control circuitry for the circuits themselves will likely be conventionally designed. At present, HP believes it can deliver storage densities of approximately 12GB/cm2
on 15nm process technology with four memory cell layers stacked on top of each other. The scientists think we may see devices
as early as 2013, though that's not an official company announcement.
Companies often ludicrously underestimate time-to-market; commercial wide-screen OLED televisions have been "just a few years away" for over a decade. ReRAM may break that trend. Although neither technology is expected to bottom out by 2013, both DRAM and Flash face significant scaling problems as process technology shrinks. It's not clear that Intel's new 3D manufacturing method will offset the issues both technologies face. Memristors—even relatively slow, low-density designs—may be ready to provide equivalent performance with promises of future scaling by the time conventional techniques are exhausted.