Lupin: Spatial Resource Stealing with Outlier-First Encoding for Mixed-Precision LLM Acceleration

  • Kim, Taein; 
  • Han, Sukhyun; 
  • Kim, Seongwook; 
  • Byeon, Gwangeun; 
  • Lee, Jungmin; 
  • ... Hong, Seokin
Citations

SCOPUS

0

초록

LLM inference often exceeds on-chip memory capacity, causing frequent external memory access. Quantization reduces memory cost but loses accuracy due to outliers. Prior mixed-precision accelerators address this issue with encoding schemes, but often result in accuracy degradation for LLMs and pipeline stalls. We present Lupin, an algorithm-architecture co-design with Outlier-First Encoding, which stores outliers in high precision by reallocating less critical normal values. This preserves maximal outlier representation and enables stall-free execution with low-precision MAC units. Experiments show that Lupin maintains accuracy while achieving a 2.02× speedup.

키워드

Large language models; Mixed-precision; Outlier; Quantization
제목
Lupin: Spatial Resource Stealing with Outlier-First Encoding for Mixed-Precision LLM Acceleration
저자
Kim, Taein; Han, Sukhyun; Kim, Seongwook; Byeon, Gwangeun; Lee, Jungmin; Hong, Seokin
DOI
10.23919/DATE69613.2026.11539287
발행일
2026
유형
Conference Paper
저널명
Proceedings -Design, Automation and Test in Europe, DATE