[K-5] Advanced Neuron and AI Systems

[K-5-03] Neutron-induced stuck error bits and their recovery in DRAMs on GPU cards

〇Masanori Hashimoto¹, Yangchao Zhang², Kojiro Ito² (1. Kyoto Univ. (Japan), 2. Osaka Univ. (Japan))

Presentation style: On-site (in-person)

https://doi.org/10.7567/SSDM.2022.K-5-03

Although soft error occurs randomly in neutron-irradiated DRAM, some bits repeatedly cause multiple errors during and even after neutron irradiation. Such a stuck bit error has been reported in the literature. This work conducted several error-checking experiments for DRAM on GPU cards to understand this phenomenon thoroughly. We observed stuck bit errors remaining for months. Also, we found that stuck block error could occur and revealed its temporal behavior of address shifting.