Open-source ML accelerator on Zynq SoC shows radiation sensitivity

Researchers have characterized the proton irradiation response of an open-source neural network accelerator deployed on a Zynq UltraScale+ SoC. The study subjected the system to 20 to 58 MeV proton irradiation, delivering a total dose of $4.29 imes 10^{10}$ p/cm$^2$. During operation, the system experienced seven workload interruptions, including restarts, reboots, and a power cycle. Two instances of output corruption were observed, where the accelerator incorrectly classified CIFAR-10 images, with one event returning a class not present in the dataset for 39 consecutive inputs. These findings highlight the need for robust software hardening and end-to-end content checks for COTS FPGA-SoCs used in space systems for neural network inference. AI

IMPACT This research provides crucial data for developing more reliable AI hardware for space applications, potentially enabling wider adoption of ML in harsh environments.

RANK_REASON The cluster contains an academic paper detailing experimental results on hardware resilience. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CIFAR-10
• FPGA-SoCs
• Linux
• LPDDR4
• ResNet-20
• Zynq UltraScale+ MPSoC

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →