Editorial: In-memory sensing and computing: New materials and devices meet new challenges

Changjin Wan, Zhongrui Wang, Rohit Abraham JohnFrontiers in Nanotechnology 2022 · DOI · OpenAlex

Introduction

The convergence of Artificial Intelligence (AI) and the Internet of Things (IoT) is fundamentally redefining industrial and economic operations. As these technologies evolve, the demand for energy-efficient, high-performance computing at the "edge"—where data is generated—has increased exponentially. Traditional computing paradigms are increasingly ill-equipped to handle the real-time processing requirements of sophisticated sensor and actuator systems used in robotics and bioelectronics. This editorial explores the shift toward neuromorphic computing and the emerging field of in-memory sensing.

The Challenge: Overcoming the von Neumann Bottleneck

Conventional computer architectures are defined by a physical separation between the processing unit and the memory unit. This structure leads to the "von Neumann bottleneck," where the constant movement of data between these two units consumes significant time and energy. For edge devices like wearable health monitors or autonomous drones, this energy overhead is a critical limitation.

Furthermore, current digital systems struggle when interfaced with real-world sensors. Sensory inputs are typically non-structural, non-normalized, and fragmented. Translating this "messy" data into a format suitable for traditional digital processing incurs massive energy, time, and wiring overheads. This calls for a paradigm shift where sensing and processing are no longer distinct operations.

The Solution: In-Memory Sensing and Computing

Inspired by the biological brain's parallel processing and low power consumption, neuromorphic computing offers a way forward. By performing computations directly within memory, this paradigm eliminates the data transfer bottleneck, improving computational capability while providing additional area and power savings.

The next frontier is "in-memory sensing," which fuses sensing, memory, and processing capabilities into a single physical unit. This mimics biological systems, such as the human eye or skin, which process information locally before transmitting signals to the brain. This fusion is essential for unleashing the full potential of advanced bioelectronics and robotics, allowing for "bio-plausible" systems that handle fragmented sensory data much more gracefully than digital systems.

Key Research Directions and Methodologies

This Research Topic highlights several critical methodologies driving the field:

  • Logic-in-Memory (LIM): Researchers are utilizing emerging Non-Volatile Memory (NVM) devices—including OxRAM, CBRAM, PCM, and FeRAM—to perform hardware-based operations. A notable application is the implementation of cryptographic functions, such as the Keccak-f hash function, directly in hardware to secure IoT devices efficiently.
  • Flexible and Wearable Electronics: The development of memristive arrays using inorganic, organic, and 2D materials is paving the way for "electronic skin" (e-skin). These materials allow for devices that can be stretched or bent while maintaining computational performance.
  • Architectural Trade-offs: The field is currently evaluating two primary device structures:
  • Two-terminal memristors: These offer high integration density and a small footprint, ideal for large-scale synaptic arrays.
  • Multi-terminal transistor-based bionic devices: These provide superior parallel processing capabilities and complex signal modulation, better suited for multi-input sensing tasks.

Key Insights and Performance Gains

The shift toward hardware-level fusion provides significant advantages over traditional software-based approaches:

  1. Massive Efficiency Gains: Hardware-based mapping of algorithms can lead to dramatic performance leaps. For instance, research by Kingra et al. demonstrated a 300-fold improvement in the Energy Delay Product (EDP) for hashing functions using NVM devices compared to conventional software-based strategies.
  2. Hardware-Level Security: Integrating security algorithms like hashing directly into memory hardware drastically reduces the energy cost of cybersecurity for resource-constrained IoT devices.
  3. Material Diversity: The field is moving beyond silicon, exploring 2D materials and organic compounds to achieve the flexibility and biocompatibility needed for integrated electronics.

Limitations and the Path Forward

Despite its promise, in-memory sensing and computing is still in its infancy. Several hurdles remain:

  • Implementation Gap: While theoretical concepts and individual devices are advanced, creating reliable, mass-producible hardware that integrates seamlessly with existing software ecosystems is a significant challenge.
  • Interdisciplinary Collaboration: Progress requires a tight feedback loop between materials scientists (developing new compounds), microelectronics engineers (designing circuits), and computer scientists (optimizing algorithms).
  • Scaling: The primary goal for the next decade is moving from isolated bionic devices to large-scale, integrated neuromorphic systems capable of navigating complex, real-world environments.

As hardware finally begins to catch up with advanced software simulations, the realization of truly autonomous, energy-efficient edge AI is becoming a tangible reality.

Summary updated

Abstract

With the convergence of Artificial Intelligence (AI) and Internet of Things (IoT) redefining the way industries, business, and economies function, the demand for energyefficient and high-performance computing at the edge is exponentially increasing. Inspired by the low power and parallel processing capabilities of the biological brain, Neuromorphic Computing is an emerging computing paradigm that overcomes many limitations of the conventional computer architecture. Most importantly, by performing computations in-memory, Neuromorphic Computing overcomes the von Neuman bottleneck, thus improving the computational capability along with additional area and power savings. While several stand-alone neuromorphic chips have been developed with excellent energy efficiency for running specific AI algorithms, such digital systems still suffer when interfaced with edge sensors. This is because the sensory inputs are nonstructural, non-normalized, and fragmented, which incur large energy, time and wiring overheads on digital systems with separated sensing and processing units. This calls for in-memory sensing technologies, with fused sensing, memory, and processing capabilities, to unleash the full potential of highly sophisticated sensor and actuator systems used in bioelectronics and robotics.
BibTeX
@inproceedings{wan2022editorial,
  author = {Changjin Wan and Zhongrui Wang and Rohit Abraham John},
  title = {{Editorial: In-memory sensing and computing: New materials and devices meet new challenges}},
  booktitle = {Frontiers in Nanotechnology},
  year = {2022},
  doi = {10.3389/fnano.2022.1073863},
}
Figure