Bullets and Roses: Memristors

Pratima H

INDIA: Doesn’t matter whether it is the wrist-work of the driver of a wooden cart with wobbly wheels or a V8 roaring inside a Lamborghini; engines did and will always matter.

Internally-tucked guts operate with the same significance for computing roads, and only their names keep changing.

You may know them as transistors, RAM, DRAM, NAND, Flash, in-memory or Memristors; what they boil down to is the engine that powers or breaks a jalopy or a race-horse with equal force.

If you are still stuck at the word Memristors in that sentence, you couldn’t have crumbled a better cookie today.

Today, we are talking about them, in all their glory and gluten. And trying to answer the same question: Are they the new V12 for our machines?

But first, some courtesy of introductions.

The Fourth Missing Piece

Usual explanations about memristors start with the words ‘a fourth class of electrical circuit’. After all, they glue together the square made of the resistor, the capacitor, and the inductor. This extra component ushered out of closets when in 1971 nonlinear circuit theorist Leon Chua proposed them in a new way. As an electrical component capable of constraining and regulating the flow of electrical current in a circuit, they were interesting. But what made them really magnetic was this – they could remember the amount of charge that has previously flowed through it, through the voltage element.

This made them quite distinct to transistors that use a flow of electrons, but memristors brought together electrons with ions; and hence they could retain the data even when the power was turned off.

This did not happen with a transistor, which lost information once the flow of electrons was interrupted with the unplugging of power. But memristors, no – they were non-volatile and that became their neon placard – memory augmenting a transistor.

The resistance changed with the flow inside but irrespective of expansion and contraction of resistance inside this device, their unique strength of storing resistance values as a nonvolatile memory brought them into a sharp spotlight.

Also, in a way they made the world of physics go through a deer-in-the-headlights moment, challenging the fundamental relationship in passive circuitry assumed to be between voltage and charge. Now they brought in the idea of flux, or charge or changes-in-voltage.

As per some definitions: "a memristor's resistance varies according to a device's memristance function, allowing, via tiny read charges, access to a 'history' of applied voltage."

With this, they made hysteresis, usually an anomaly in conventional circuitry more of a quintessential aspect of passive circuitry. So they can easily put on clothes of electrical devices with memory; as resistance varies according to the dynamic evolution of internal state variables.

But what does all this gobbledygook mean in actual world out here?

Computing not just redefined, but ripped apart

When these devices became the talk of the town, there were bets on the end of the electronics-age in favor of the advent of an era of ionics, and in short – the end of binary code.

These devices could remember the amount of charge that was flowing through the material and retain the data even when the power is turned off. That meant they could recall not just 1s or 0s but fractions in between too.

That was just one feather in the bow. As experts peeked into this marvel more, they found out possibilities of way-faster speeds, fairly lower power consumptions, and amazing densities of information storage compared to silicon microchip transistors.

If one could have a device that can remember voltage, or the information permanently, can switch in nanoseconds, behaves well on the power side and does not disappoint on the density side, wouldn’t one swipe away DRAM, flash, and disk away from the table?

There were all sorts of possibilities being baked – of super-fast memory chips which pack more data and gobble less energy; of more powerful computers, and ultimately of marrying two extreme-personalities together: compute and memory.

Because if memory stays even when the machine is turned off, if a new architecture can help in multiple crossbar memories stacked together, if memory could also put another hat and function like a CPU; then the world was just opening its eyes to a new eon.

The whole idea of dynamically changing memristors between memory and logic operations was a seismic enough computing paradigm. With a memristor’s ability to work out material implication logic, interconnected to create any logical operation, so much was possible.

And required too, since despite its long-running marathon all these years, Moore’s law was facing skepticism and the need for re-inventing computing processing and power was gaining more ferocity. Plus, the notion that these magical pixies could mimic neurons, self heal their weaknesses and learn fast, analogous to a human brain’s synapses and axons; there were many who started jumping the gun on a new accomplice to AI too with the arrival of memristors.

The device’s basic property is that its electrical resistance can be changed by voltage application and that brings in a memory of the voltage which is what would encode data. To add to that, memristors apparently would not need a silicon layer and hence would open doors to a new genre of microchips, that would be easily knitted into our every day window-panes, T-shirts and tea-mugs.

So was this all ‘whats’ and ‘ifs’? Did anyone bet on memristors in a serious way?

Piece De Resistance

Turns out that memristors found a lot of big players in their corner soon. Apart from the big project tagged ‘The Machine’, HP has been one of the most fluorescent pioneers for chasing memristors in full throttle ways. This includes the first stable prototype and work on titanium di-oxide. Alongside, there are other notable names like IBM, HRL, Samsung, Nantero, Crossbar and a few research labs who are fiddling passionately with the concept.

Yet, like every disrupting wave, this one is taking time and ebbs to whip up its real future.

For one, the manufacturing costs stay not so encouraging. Then a memristor is a lot about the material which must have a resistance capable of reversibly changing with voltage. From titanium dioxide, optical fibre, to molybdenum disulfide nanosheets; the work on that is always in motion.

Another aspect worth noting is that even if an ideal memristor is supposedly symmetrical to enable an even and smooth relationship between current and voltage; real-world memristors incline towards lopsided current-voltage characteristics.

What to do with them - that is still a big question out there flitting about in a wayward manner. Should they be used for something because they are very fast to switch but operate probabilistically? Or should they be employed for self-learning muscles? Or for their capacitance (with memory)?

Or for use where brain-like plasticity is better suited? Is non-volatile memory the least interesting thing about memristors? Now that’s something that experts like Todd Hoff rightly wonder.

Are we lacking in the kind of applications and algorithms that will tap the actual marrow of memristors?

Even as those questions swim about, there are doubts over how much system redesign these devices will entail, or how plug-and-play would their integration in existing infrastructures be? Plus, making chips that fit existing production lines of silicon variety is a huge question too. Another one is that of performance bottlenecks being in the interconnects, which HP is attempting to nail through photonic interconnects.

There are many concerns that will have to be addressed before memristors etch computing industry’s memory as expected – around reliability; the Landauer’s principle of the minimum energy costs; error chances, ease of programming; clock cycles; latency characteristics, thermodynamic questions, and density issues. Would large arrays perform the way individual memristors work?

To boil it down into one tea-bag, how soon would it take before memristors could elbow out DRAM and SRAM, Flash, FPGA, ASICs and the whole enchilada of memory and storage as of now?

Would the one-atom-thick nanomaterial, graphene be the missing piece here? Would 3D Xpoint memory work by Intel and Micron be the direction we want or would it take The Machine from HP to suss out memristors’ actual future?

For those answers to accelerate, memristors would have to travel way down – in other words - Under the hood.