The Potential of Memristors for Computing in 6G and Beyond

Memristors are resistors with memory. They promise high levels of integration, stable non-volatile resistance states, fast resistance switching, and excellent energy efficiency—all very desirable properties for next generation of memory technologies.

This short video is a great introduction to memristors.

The concept of memristor was introduced over 50 years ago. In 1971, a scientist named Leon Chua claimed that the field of electronics was missing something fundamental: a resistor that remembers how much charge flows through it and changes its resistance accordingly. He named this missing circuit component the memristor. They were nearly forgotten for almost four decades. 

There are many different flavors of memristive technologies. Still, in their most popular implementation, memristors are simple two-terminal devices with the extraordinary property that their resistance depends on their history of electrical stimuli. In other words, memristors are resistors with memory. They promise high levels of integration, stable non-volatile resistance states, fast resistance switching, and excellent energy efficiency—all very desirable properties for next generation of memory technologies.

The landscape of memristor-based systems for AI. In-memory computing aims to eliminate the von-Neumann bottleneck by implementing compute directly within the memory. DL accelerators based on memristive crossbars are used to implement vector-matrix multiplication directly using Ohm's and Kirchhoff's laws. SNNs, a type of ANNs, are biologically more plausible and do not operate with continuous signals, but use spikes to process and transfer data. Memristor systems could provide a hardware platform to implement spike-based learning and inference. More complex functionalities (neuromorphic), beyond simple digital switching CMOS paradigm, directly implemented in memristive hardware primitives, might fuel the next wave of higher cognitive systems.

MIT's faculty Jennifer Rupp and Ericsson's Dr. Saeed Bastani sat in an interview with All About Circuits (AAC) on the topic of Lithium-based memristors for neuromorphic computing recently:

As it turns out, the road to neuromorphic computing will hinge heavily on materials science. Lithionics, a new field emerging from MIT, investigates how lithium oxides may be such a material. Additionally, researchers from both MIT and Ericsson are exploring how memristors, or memory resistors, based on lithium oxide films may be the key to creating computer chips that mimic the human brain.

In data published in Nature Review Materials and Advanced Materials, MIT Professors Rupp and Bazant and their teams evaluated two different lithium-based films against several criteria, including:

  • Conductance
  • I–V cycling response
  • Resistance retention time of the films under low bias

The results indicated that one of the two films, Li7Ti5O12, would be appropriate for deep neural networks (DNNs) while Li4Ti5O12 was more suited to spiking neural networks (SNNs) applications.

Dr. Rupp and her team intend to explore just how challenging it might be to integrate lithium-based computing hardware with existing silicon-based technology. "There’s this existing technology [silicon], that is very far developed and you don’t want to change that too much," she comments.

Dr. Rupp also mentions another potential spin-out from lithionics research—the possibility of “micro-batteries” on silicon down to the microchip level.

Read the complete article here.

There is a lot of potential of memristors but it remains to be seen if they are ready in time for 6G to make an impact. Future application will involve smartphones, wearables and even companion gadgets. 

Posted by on

Comments

Post a Comment