Exploring subthreshold processing for next-generation TinyAI.

The energy demands of modern AI systems have reached unprecedented levels, driven by the rapid scaling of deep learning models, including large language models, and the inefficiencies of current computational architectures. In contrast, biological neural systems operate with remarkable energy efficiency, achieving complex computations while consuming orders of magnitude less power. A key mechanism enabling this efficiency is subthreshold processing, where neurons perform computations through graded, continuous signals below the spiking threshold, reducing energy costs. Despite its significance in biological systems, subthreshold processing remains largely overlooked in AI design. This perspective explores how principles of subthreshold dynamics can inspire the design of novel AI architectures and computational methods as a step toward advancing TinyAI. We propose pathways such as algorithmic analogs of subthreshold integration, including graded activation functions, dendritic-inspired hierarchical processing, and hybrid analog-digital systems to emulate the energy-efficient operations of biological neurons. We further explore neuromorphic and compute-in-memory hardware platforms that could support these operations, and propose a design stack aligned with the efficiency and adaptability of the brain. By integrating subthreshold dynamics into AI architecture, this work provides a roadmap toward sustainable, responsive, and accessible intelligence for resource-constrained environments.