AiF: Accelerating On-Device LLM Inference Using In-Flash Processing

초록

While large language models (LLMs) achieve remarkable performance across diverse application domains, their substantial memory demands present challenges, especially on personal devices with limited DRAM capacity. Recent LLM inference engines have introduced SSD offloading for model parameters to reduce memory footprint. However, the highly memory-bound nature of on-device LLMs makes inference speed heavily dependent on read bandwidth, leading to significant performance degradation due to the limited bandwidth of SSDs. In this paper, we propose an in-flash processing solution for on-device LLM, called Accelerator-in-Flash (AiF), which integrates matrix-vector multiplication (GEMV) operations directly into flash chips. By enabling in-flash GEMV operations, AiF leverages the high internal bandwidth of flash chips without being constrained by the limited external bandwidth. Building on this core structure, AiF employs two novel flash read techniques that were specifically optimized for reading LLM parameters stored in flash memory. AiF achieves a 4x boost in internal read bandwidth during inference with minimal implementation overhead, thanks to its streamlined error correction process. Evaluations on eight real-world LLMs reveal that AiF provides a 14.6x throughput improvement compared to baseline SSD offloading schemes. Furthermore, AiF surpasses in-memory inference, delivering 1.4x higher throughput with a significantly reduced memory footprint.

키워드