Integrated circuit architectures for time-based intelligent sensor interfaces

Summarise with VIRAT
Authors: Ahmed M. Mohey Attalla
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Smart sensory systems have been increasingly adopted in our daily lives, driven by numerous applications and use cases, with billions of sensor nodes serving this purpose. A smart sensor node is typically equipped with interface, communication, and processing (named "in-sensor computing") functions. To realize these functions within a tiny node, the sensor interface must be efficient, compact, and scalable. Moreover, the integration of the interface with the other node's components, such as digital processing circuits, should also be eased, promoting digital-friendly architectures. Insensor computing blocks undergo similar constraints. They must have low energy consumption and a compact form factor. Traditional analog architectures may suffer from performance limitations in deep sub-micron technologies due to the reduced voltage headroom/gain, leakage, and mismatch. As an alternative, this thesis explores time-based architectures to realize scaling-friendly and efficient circuits for smart sensor nodes. Specifically, this work presents several time-based blocks for sensing applications: (1) time-based sensor interfaces, with a focus on time-to-digital converters (TDCs), which are core building blocks in these interfaces; (2) time-domain neural network accelerators, which are the main components to realize in-sensor computing. First, TDC architectures are proposed to target a broad range of low-power and event-triggered sensing applications, such as time-of-flight (ToF) sensors. They employ gated ring oscillators (GROs), all-digital control, and all-digital readout, achieving competitive performance metrics in terms of scalability, dynamic range, power dissipation, and immunity against sampling errors. The proposed TDCs achieve a dynamic range up to 1.6ms, while dissipating an average power as low as 39.37μW in the active mode of operation and only a few hundred nanowatts in the idle mode of operation. Second, a time-domain neural network architecture is proposed to target precision insensor processing for applications like human activity recognition (HAR). It achieves good performance metrics thanks to its high precision, small size, and low complexity. Moreover, it has an all-digital implementation, allowing easy tailoring of the design to suit various sensing applications with no dependency on analog non-idealities. The proposed time-domain neural network accelerator supports 8-bit inputs/weights and 18-bit accumulation (outputs) in a 612.5μm2 area per MAC, which is considered among the smallest implementations.

Fields of Study & Subjects

Analog and Mixed-Signal Circuit Design Advancements in PLL and VCO Technologies Low-power high-performance VLSI design

Publication details

Journal / source
Aaltodoc (Aalto University)
Publication date
1 January 2026
Language
EN
Document type
Thesis