Tech by VICE

New Technology May Beat the 'Memory Bottleneck' with Plain Old Silicon

Fast computers need fast data. This is a problem.

by Michael Byrne
Apr 20 2016, 1:43am

Silicon wafer. Image: Wikimedia

Among the holy grails of computer engineering is a concept/technology called non-volatile universal memory. Its promise is nothing less than the wholesale replacement of every one of the disparate memory technologies that computing currently depends upon: DRAM, SRAM, SSD, hard drives, etc. In their stead would be a single unified technology that fills all of the memory requirements of a computer—that is, it will be ultrafast, infinitesimally small, and non-volatile, e.g. it will be able to retain data even when powered down, like a hard-disc or SSD but with RAM speeds.

This is more than a matter of convenience. As computers become ever faster and more powerful, they are increasingly subject to what's known as the "memory bottleneck." Simply put, computers are processing data faster than the hardware architecture can deliver that data. This will only become more of a problem.

Materials scientists from Moscow Institute of Physics and Technology, the University of Nebraska, and the University of Lausanne in Switzerland have a new solution in the form of a nanoscale polycrystalline ferroelectric film. While ultra-thin ferroelectric materials are hardly a novel target for universal memory, the new scheme, which is described in the journal ACS Applied Materials & Interfaces, adds the highly desirable feature of silicon integration. Meaning, it's theoretically possible to fabricate the ferroelectric films in existing semiconductor manufacturing facilities (with some modifications).

Image: Zenkevich et al

The ferroelectric memory idea is simple enough. It's roughly analogous to what occurs in existing DRAM memory, in which data is stored in individual memory cells consisting of a capacitor and a transistor. In the DRAM cell, the presence or absence of an electrical charge within the capacitor represents a binary bit of information. These bits are read and written via the transistor, which either adds charge to/drains charge from the capacitor.

The catch with DRAM is that capacitors will leak charge over time. This necessitates a continuous refreshing of a DRAM memory unit—leave it powered down, and eventually the DRAM cell will lose its information content.

Thin-film ferroelectric memory replaces the capacitors of DRAM with a variation that's based on magnetic polarization. Instead of charge or no-charge, we have magnetized up or down. Not only does flipping polarities require much less energy, polarization doesn't require the continual refreshing of standard DRAM memory.

The team behind the current paper managed to get their ferroelectric material down to a thinness of 2.5 nanometers, which is about one-tenth the size of the smallest virus. The underlying concept is known as a ferroelectric tunnel junction. It consists of the film itself sandwiched between two tiny electrodes which work together to change the polarization of the film, or the ultratiny cell-speck of film in between the electrodes. Information is then read from the film by probing the electrical resistance of the cell, which will change according to its polarity.

And there you have it: information storage and retrieval at atomic scales.

Here, the ferroelectric film has been "grown" on top of a silicon substrate via a process called atomic layer deposition, which is about what it sounds like. The resulting 2.5 nm material (hafnium oxide) was tested using several techniques confirming that it indeed has the desired ferroelectric properties, e.g. is both stable and switchable. "To the best of our knowledge, so far no ferroelectric behavior has been reported for thinner [hafnium oxide]-based films," the researchers note in the study.

"The work to demonstrate the so called tunneling electroresistance effect in a prototypic memory device is under way now," Andrei Zenkevich, the study's lead author, told EE Times. "Judging from pulsed measurements of the polarization reversal, the prospective write time is within the nanosecond range. The reading of the information occurs non-destructively by measuring the (tunneling) current through the junction and access time should mainly depend on the electronics circuitry."

Don't go and dump your SSD quite yet. Even just confirming all of this will take years. Still, it seems clear enough that the memory problem has a solution—if not this, than one of any number of alternative avenues being pursued for nanoscale non-volatile memory at this very moment.