Silicon Valley is mostly a world of practical technology—applying principles from pure science to create handy gadgets. But today, Hewlett Packard announced a new electrical component born of theoretical physics. The device, a nanoscale component called a “memristor,” requires no power to retain data, which it can store more densely than a hard drive and access about as fast as a computer’s RAM memory—potentially allowing it to replace both components in the future.
Memristors can function in either a digital mode, in which a memory cell is “on” or “off,” or in analog mode, in which each cell holds some value in between. These values grow every time the cell receives an electrical signal, mimicking the way neurons in the brain build stronger memories the more they are stimulated.
The memristor was theorized in the early seventies by an electrical engineer name Leon Chua, but it took decades before anyone could prove it exists. The process was similar to particle physics, in which mathematicians first propose a particle and then experimenters eventually find it—or don’t. (The hottest current example is the search for the theoretical Higgs boson, aka the “god particle.”) Chua’s mythical electrical component didn’t show up until recently when HP researchers were studying the electrical properties of nanoscale materials and came upon a few that acted suspiciously like the memristor. After some refinements, they invented exactly what Chua theorized.
It’s a long path between proving something in the lab and selling it at Best Buy. Stan Williams, who leads the HP research team, expects memristors to first show up in the next few years as “cache” that sits between a hard drive and the DRAM memory in PCs. The hard drive could load key data, like the instructions to start up Windows, into the memristor cache, which can dump it into the DRAM far faster than transferring it straight from the hard drive—resulting in lightening-fast boot-ups and quick opening of large files.
But Williams has bigger plans to eventually replace both the hard drive and the RAM with one memory system that eliminates the need to store data on a relatively slow hard drive and then laboriously load it into fast DRAM before the PC can use it. He thinks a memristor can hold scads more data than a hard drive and access about as fast as DRAM. “You expand out both ways and try to eat the heart out of both the DRAM and the hard disk,” said Williams.
Williams also wants to eat into the CPU. He says that many processes requiring fuzzy logic, like recognizing a face, are very hard to work out with that yes/no logic of a digital computer. But for an analog computer—like, say, the brain—it’s a piece of cake. So Williams proposes a CPU with multiple processing cores: Some digital for the number crunching that today’s computers do so well, and others using analog memristors. Take facial recognition. Someone’s face can change from day to day, and definitely from year to year. But it’s not a drastic change that a yes/no digital system is good at figuring out. It’s a mild difference with some degree of change along a continuum that Williams says is perfect for an analog computer.
Could Williams take it even further—creating not just a PC that thinks a little more like a human, but an electronic mind that thinks exactly like one? “We’re not claiming we’re going to build a brain anytime in the next decade,” he says. But he’s not ruling it out for later on.
How it Works
The unlikely invention of the memristor took about 40 years. It started in the early 1970s when electrical engineer Leon Chua was looking at the interplay of electrical forces in the basic elements of circuits: resistors, capacitors and inductors. The same math that explained those three elements indicated that there should be a fourth, which he named the memristor (short for “memory resistor”) in a 1971 paper. A memristor would change its level of electrical resistance if charge were applied and retain or “remember” that resistance until another charge were applied. Like the Higgs boson “god particle,” the memristor made perfect sense on paper, but no one had ever seen one.
Not until the late 1990s, when researchers at Hewlett Packard Labs were studying the electrical properties of different nanotech materials and found several that looked pretty similar to Chua’s hypothetical memristor. Suspecting that Chua’s mythical circuit was real, HP researchers set out to invent one.
What they wound up with was deceptively simple: two layers of a semiconductor, titanium dioxide, sandwiched between electrodes. The bottom layer contains the standard material, which is virtually useless for conducting electricity. The top layer is missing a few oxygen atoms, creating positively charged “bubbles” that make it a conductor.
Running a positive charge through the electrode above this layer pushes some of the charged bubbles into the lower layer (where they stay, until another charge is applied), allowing it to conduct electricity and lowering the electrical resistance of the entire cell. A computer can read information in a memristor cell by measuring how much resistance it has.
Hitting the cell with relatively big zaps switches it from high to low resistance levels that correspond to zero and one for digital data. Using less power can give it some resistance value in-between the extremes. And the results are cumulative. The more often you charge the cell, the lower its resistance—in other words, the stronger its memory becomes. That matches the way neurons build stronger connections over time to make memories stronger (and explains why the more you practice something like piano, the better you get at doing it). Applying a negative charge to the top of the cell reverses the process—a lot of power switches a one to a zero in a digital system. Applying finer amounts of current causes a memory to fade.
HP’s memristors are tiny, about 15 nanometers across. That allows them to store data about as densely as a hard drive—100 gigabits per square centimeter. But HP thinks it can get them far smaller—down to four or even two nanometers. Even at 4nm, a square centimeter of memristor can hold one terabit.
The complete explanation is in a paper that HP published this morning in the journal Nature.