상세 보기
초록
This paper presents a multi-threaded frontend ar-chitecture for neural processing unit (NPU), called Function Unit (FU)-aware Simultaneous Multi-threading (SMT)-like. Each thread operates on an independent pipeline, with arbitration used only when contention arises for shared FU controllers. Dependency tracking is handled per thread. Our design reduces latency by an average of 3.7x and improves functional unit utilization by 3.4x compared to a single-threaded baseline.
키워드
Frontend; FU-Aware SMT-Like; NPU
- 제목
- A Function Unit-Aware Frontend Architecture for Multi-Threaded Neural Processing Unit
- 저자
- Kong, Jaeyoung; Park, Jeongwoo
- 발행일
- 2025
- 유형
- Conference Paper
- 저널명
- International SoC Design Conference 2025, ISOCC 2025 - Proceedings of Technical Papers