講演情報
[14p-P10-13]A Via-Programmable Neuron Array for Wearable AI-IoT Applications
〇Jaewon Shin1, Rei Sumikawa1, Dongzhu Li1, Mototsugu Hamada1, Atsutake Kosuge1 (1.UTokyo)
キーワード：
via-programmable neuron array、wearable AI-IoT、cost-efficient hardware
This study proposes a low-cost, high-efficiency via-programmable neuron array (VPNA) for wearable AI-IoT devices. By using a single programmable via mask while keeping other layers fixed, VPNA reduces manufacturing costs. Key techniques, including the bit- and neuron-serial circuit (BNSC) and the function-selective non-linear neural network (FS-NNN), optimize wiring and resources. BNSC reduces wiring complexity by 1,024 times through bit-serial communication and a multi-drop bus configuration, minimizing power and area. FS-NNN simplifies activation functions, limiting them to 4 types: ReLU, inverse ReLU, linear, and inverse, decreasing hardware requirements while ensuring flexibility. Combining these techniques, VPNA provides a cost-effective, scalable, and energy-efficient solution for next-generation AI-IoT applications.