A FUTURE RE-IMAGINED

Dick Pountain/PC Pro/Idealog 232  06/11/2013

Once upon a time, in the '80s and early '90s, I used to write highly technical columns about future computer technologies, both hardware and software - stuff like asynchronous or transport-triggered processor architectures, object-oriented memory managers, parallel processing algorithms and much more. I don't do that stuff much nowadays, but the reason is not that my interest has waned but rather that the total triumph of Intel and Microsoft condemned many of those esoteric research avenues to become dead ends. And my aging brain rebels when asked to study the detailed cache architecture of Intel's next CPU generation. That doesn't mean I've lost interest entirely, merely that I can now afford to wait for breakthroughs that might just change the whole game, which don't happen very often (and don't always deliver). Over the last 10 years I've covered just three such technologies, namely spintronics, diamond-film quantum dots and graphenes. It still gives me a bit of a thrill when a new one arrives, as it just has with the "memristor".

If you've heard of the concept it will almost certainly be thanks to Hewlett Packard's occasional announcements that it's working on memristor-based storage devices, and hopes to have 100Terabyte drives available in around five years. But memristors, if they can be made to work economically - which is not yet certain - promise more than storage. If they are able to function like transistors too they could render possible the first wholly-integrated computer architectures in which CPU, local memory and mass storage are all built from very similar basic units.

So what exactly is a memristor? As its name suggests, it's an electronic component that combines the attributes of a resistor and a memory. What that means in HP's application is a kind of resistive RAM, in which the memory cells are resistors that retain a memory of the current that last flowed through them. Pass a current through the cell one way and its resistance increases, pass a current the opposite way and resistance decreases, but crucially the cell *preserves its last state* and so acts as a non-volatile memory. You can read 0s and 1s by measuring the resistance of cells. HP's memristor cells are implemented by two sets of parallel wires, one platinum and one titanium, at right angles and separated by twin layers of titanium dioxide. One layer is pure oxide and a good insulator, while the other is lightly doped to deplete it of some oxygen atoms. Where they cross a current passed through the junction causes oxygen "holes" to migrate into the undoped layer, reducing the resistance of both layers: when the current is reversed they migrate back again. Two things to note about this scheme: the substrate doesn't need to be silicon, and the cells are so simple they can be made very small indeed, giving huge packing densities.

HP hopes to make 3nm memristors that switch in a nanosecond, as fast as DRAM and denser than flash memory, which they could replace. All very good, but the memristor concept goes deeper. The Chinese-American non-linear circuit theorist Leon O. Chua has suggested that the memristor is an expected fourth fundamental electronic device - the better-known ones being the resistor, capacitor and inductor - that links magnetic flux to electric charge. It should be able to do more than just make memory cells. Other researchers have shown that memristors can be made using different physical effects, like magnetic spin or polarisation, and from different materials like organic polymers. It's also been suggested that memristor cells can be combined into "crossbar latches", a novel kind of logic gate that can function like a transistor in a processor architecture. Most suggestive of all is that memristors remember in a way similar to the reinforcement of synaptic connections in the animal brain, and there are teams working on using them in neural learning networks.

Bundling all such speculations together sets my imagination running riot. Throw in a few other leading-edge technologies like graphenes with their miraculous conductivity and photosensitivity, micropipelines for asynchronous operation, and analog resistive networks as used by Carver Mead for his chip that emulates the human retina. Now imagine future flexible plastic materials that can collect solar power and store it in their own integral hypercapacitors, that contain petabytes of memory integrated with distributed processing elements, that can see using grids of nano-lenses that mimic the insect compound eye, can move using grids of nano-actuators and can learn using integrated neural networks. Sci-fi authors have forseen such materials in their imagined futures for many, many years, and though we may be nowhere near achieving them, we're beginning to see dimly where the route might start. Perhaps memristors are a crucial step onto that route.