Design of a nanosensor array architecture

Vijaykrishnan Narayanan, Wei Xu, Narayanan Vijaykrishnan, Yuan Xie, Mary Jane IrwinACM Great Lakes Symposium on Vlsi 2004 · DOI

TL;DR

This paper introduces NANSEA, a novel, compact, and fault-tolerant architecture for nanowire sensor arrays designed for high-speed electronic noses. By integrating simple processing elements (PEs) directly with sensors and employing a minimum distance classifier, NANSEA aims to overcome the limitations of traditional, power-hungry, and less scalable sensor systems.

What Problem Does This Paper Address?

The development of electronic noses, which mimic biological olfaction for detecting and classifying odors, faces significant challenges. Traditional approaches often rely on large arrays of sensors coupled with off-chip pattern recognition, leading to high power consumption, large chip areas, and scalability issues, especially when dealing with hundreds or thousands of sensors. Furthermore, the low yield of nanowire fabrication (around 60%) necessitates a system that can gracefully handle faulty sensors. This paper addresses the need for a compact, low-power, and fault-tolerant on-chip architecture for nanowire sensor arrays that can perform both sensing and pattern recognition efficiently.

Methods

The core of the proposed solution is the NANSEA (NANnowire SEnsor Array) architecture. Key methodological contributions include:

  • Massively Parallel Architecture with Simple PEs: Each processing element (PE) is directly connected to seven sensors. These PEs are designed to be very simple, performing basic computations to minimize area and power.
  • RMOT Organization: The architecture is built upon a "reduced mesh of trees" (RMOT) organization, where PEs are placed along the diagonal of the sensor array, providing efficient access to both rows and columns of sensors. This design separates processing and sensing areas.
  • Low-Resolution A/D Conversion: To reduce complexity and power, a 1-bit resolution analog-to-digital converter is employed. This is made feasible by the significant resistance change (AR) in nanowire sensors, which is orders of magnitude larger than the base resistance.
  • Fault Tolerance through Mark Matrix: A "Mark Matrix" is used during a training phase to identify and mark faulty sensors. This allows the system to focus only on working sensors, mitigating the impact of low fabrication yields.
  • Minimum Distance Classifier: For pattern recognition, the paper implements a minimum distance classifier in software. This algorithm is chosen for its ease of implementation and lack of probabilistic assumptions, making it suitable for real-time decision-making.

Datasets

Not discussed in the paper. The paper focuses on the architecture and simulation results, not on training with specific real-world datasets.

Key Insights and Ideas

  • Compactness through PE Simplification: The realization that very simple PEs, directly interfaced with sensors, can achieve high-speed sensing and pattern recognition is crucial. This contrasts with more complex, general-purpose processors.
  • Fault Tolerance as a Design Principle: Explicitly addressing the low yield of nanowire fabrication through a "Mark Matrix" is a practical and important design consideration for real-world deployment.
  • Scalability via RMOT: The RMOT organization provides a structured way to scale the sensor array and processing elements while maintaining efficient communication pathways.
  • Trade-off in A/D Conversion: The paper demonstrates that aggressive reduction in A/D converter resolution is possible and effective when combined with sensitive sensors and appropriate pattern recognition algorithms.

Key Results

The paper presents synthesis results for the NANSEA architecture using TSMC 0.25µm and commercial 0.16µm standard cell libraries.

  • High Operating Frequency: The design achieved frequencies up to 293 MHz (TSMC 0.25µm Fast) and 289 MHz (Commercial 0.16µm), which is sufficient for high-speed sensing feedback.
  • Compact Design: The architecture is described as very compact, with a 4x4 RMOT4 block (16 PEs) implemented. The synthesis results show cell areas in the range of 108,000 to 111,000 µm² for the TSMC 0.25µm technology.
  • Low Power Consumption: Dynamic power consumption was reported in the range of 51-76 mW for the TSMC 0.25µm technology.
  • Efficient Operation: The system's operation is illustrated with an example of gas detection, showing how the Mark Matrix and minimum distance classifier work together to identify the presence of a target gas.

Limitations and Future Work

  • Software-Based Pattern Recognition: While the architecture is designed for on-chip integration, the pattern recognition algorithm is currently implemented in software. Future work aims to integrate this into a hardware engine.
  • Full Custom Design: The current implementation uses standard cell libraries. The authors are developing a full custom design with a larger array of 4x16 PEs and 16x16 nanowire sensors, aiming for further optimization.
  • Limited Dataset: The paper does not present results on extensive datasets, focusing on architectural design and simulation.
  • Specific Gas Detection: The example provided focuses on detecting the presence of "gas" rather than classifying specific types of gases, which is a common goal for electronic noses.

Summary updated

BibTeX
@inproceedings{narayanan2004design,
  author = {Vijaykrishnan Narayanan and Wei Xu and Narayanan Vijaykrishnan and Yuan Xie and Mary Jane Irwin},
  title = {{Design of a nanosensor array architecture}},
  booktitle = {ACM Great Lakes Symposium on Vlsi},
  year = {2004},
  doi = {10.1145/988952.989024},
}
Figure