부산대학교 도서관

As large language models (LLMs) expand their sizes, conventional GPU-based LLM inference systems face memory bandwidth and capacity limitations. An LLM inference accelerator using NAND flash storage has been proposed to overcome these challenges. However, this necessitates a significant expansion of flash channels to ensure adequate bandwidth for inference, subsequently escalating error correction code (ECC) costs. This paper examines the impact of flash memory errors on LLM inference accuracy and explores the possibility of lightweight ECC by leveraging LLM's inherent error resilience. We analyze the impact of 1) high-order bit indices masking for FP32 LLM parameters, 2) clipping, and 3) a dependency by parameter type of error robustness, and show that a combination of them can reduce ECC bandwidth by up to 9.38%.