Modern computers burn most of their energy on onerous data transfers. This is because of von Neumann architecture, a ubiquitous computer design that separates processors from memory units.

This design, which was first proposed way back in the forties and lauded for its flexibility and scalability, is now creating bottlenecks in the artificial intelligence (AI) era. Over the past few decades, data transfer efficiency has fallen behind improvements in processing and memory. Moreover, with AI, where most tasks are interrelated, processors often get stuck working below capacity as data shuttles back and forth. Thus, von Neumann architecture has become one of the biggest impediments to AI growth.

An NUS study published in Device now addresses this issue by introducing a new, integrated processor which performs calculations directly within memory – a method called in-memory computing. Balancing energy efficiency with high-accuracy performance, the processor is built on a reservoir computing (RC) system, which utilizes physical systems to reduce computational costs.

Neuroscience meets computing

At the core of the RC system is a two-layer architecture: a three-dimensional computational space, called the ‘reservoir’, and a second readout layer trained to read the reservoir’s output.

RC systems are an example of an emerging field of neuromorphic computing. The field borrows heavily from biology and neuroscience, drawing inspiration from the biological structure of the human brain to develop systems that operate a fraction of the operational cost of our current transistor-based computers.

The study, led by Professor Aaron Voon-Yew Thean from NUS Electrical and Computer Engineering, zeroed in on a hardware component with promising implementation potential for RC architecture, the memristor. Memristors, short for memory resistors, are a type of memory which store information based on resistance changes induced by electrical current. The physical changes of a memristor simulate the synaptic plasticity of the brain, where changes happen on the microscale in our neuronal connections as we learn and remember.

“The human brain stores and processes data locally,” Thean pointed out. “So, as information passes from neuron to neuron through the synaptic connection, part of the computation happens while it's passing through the synapse.”

Memristors, however, have thus far proved commercially elusive for RC implementations. Most of our current integrated circuits are made with complementary metal–oxide–semiconductor (CMOS) fabrication, which rely on silicon as its primary semiconductor material. But with silicon-based memristors, computer scientists struggle to capture the energy efficiency of neuromorphic features within a compact design.

Some scientists have turned to emerging materials like perovskite, but these materials suffer from high production costs and poor CMOS compatibility. Separating the two RC layers into two different materials also introduces processing and integration hurdles. Thus, for RC systems to become a practical reality, there is a clear need for a single, CMOS compatible material that can support the distinct switching behaviors required by both RC layers.

Size matters, when it comes to silicon

What Thean’s team has done differently is to reexamine the neuromorphic potential of silicon – but this time, in the form of nanosheets.

Nanosheets, they discovered, offer greater property tunability than conventional silicon wafers or silicon dioxide: arrays made from small nanosheet flakes show volatile resistive switching ideal for the reservoir, while arrays from large flakes provide nonvolatile, near‑linear conductance updates and robust memory retention suitable for the readout layer.

“There are atoms that move between the nanosheets to form a filament connecting the top and bottom electrode,” explained Thean. The properties of the filament can be controlled by the lateral surface area of the nanosheets, producing the distinct dual behaviours needed for a two-layer RC system.

The characteristics discovered helped to inform the next step, where the researchers used the arrays to build a simulated RC processor. When benchmarked against handwritten character recognition and nonlinear dynamics prediction tests, the processor achieved high accuracy.

Silicon nanosheets, the study concluded, offer strong potential for compact, energy-efficient next-gen electrical devices. They offer scalability, low processing temperatures and excellent compatibility with standard CMOS production lines.

While the memristor’s metrics are promising, the researchers highlight that improved endurance and tighter performance consistency are still needed to reach commercial readiness.

There is hope, Thean noted, that in-memory computing will not only replace parts of the modern computer, but will ultimately make technologies with high computing demands like AI sustainable well into the future.