Presentation
MCUGen: Memory Efficient Mapping and Code Generation for DNN Inference on MCUs
DescriptionDeploying DNNs to microcontroller units (MCU) is challenging because of their constrained memory resources (e.g., 128KB). In this paper, we propose MCUGen to generate memory efficient code on MCUs. The key idea is to overlap the memory space of inputs and outputs to reduce the peak memory footprint. In detail, we first use polyhedral model to minimize the peak footprint of single DNN operators analytically. Then, for the whole DNN graph, we combine the polyhedral analysis with operator fusion to enable more memory reuse. Evaluation on ARM MCUs shows that MCUGen can reduce $58.6\%$ RAM usage compared to state-of-the-art work.
Event Type
Work-in-Progress Poster
TimeTuesday, July 11th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
RISC-V
Security