14.7 NeuroPilot: A 28nm, 69.4fJ/node and 0.22ns/node, 32×32 Mimetic-Path-Searching CIM-Macro with Dynamic-Logic Pilot PE and Dual-Direction Searching
An Guo, Jingmin Zhang, Xingyu Pu, Yi Yang, Defa Wu, Yuchen Tang, Yuhui Shi, Yinghai Gao, Zhichao Liu, Bo Wang, Tianzhu Xiong, Zhaoyang Zhang, Xi Chen, Jinwu Chen, Feiran Liu, Xing Wang, Xinning Liu, Weiwei Shan, Bo Liu, Hao Cai · 2025
Autonomous micro-robots, equipped with AI for urban navigation, are being deployed for various applications, such as package delivery and surveillance. These intelligent machines require efficient path-finding algorithms to navigate complex city environments effectively. Similarly, other domains like VLSI routing, urban path planning, and wavefront propagation also demand optimal navigation solutions, as illustrated in Fig. 14.7.1. While previous work has typically relied on digital accelerators to address navigation challenges [3]–[5], mimetic-path CIM architectures, which incorporate SRAM within their processing-element (PE) arrays, have demonstrated superior performance, power efficiency, and area utilization (PPA). However, existing mimetic-CIM solutions face several challenges: (1) single-direction searching (SDS) inefficiently utilizes half of the available search time; (2) DFF-based temporary storage suffers from poor PPA, while dynamic-logic based alternatives are prone to leakage issues; and (3) limited processing windows can lead to task failures in large-scale maps, but previous attempts to address these issues, such as map direct transfer and splitting, have their own limitations. This paper presents a 28nm 32×32 mimetic-path NeuroPilot CIM macro with (1) dual-direction searching (DDS) to accelerate the path searching; (2) a dynamic-logic pilot PE with an outspread pre-charge (OPC) circuit, achieving a superior PPA while reducing leakage; and (3) a novel 3-step coarse/fine-path (TsCFP) flow enabling efficient navigation of large maps. The proposed chip achieves a search rate of 3670M nodes/s, 213.6μm2PE area, and a 69.4fJ node energy consumption.