http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
NPU를 위한 효율적인 Fused Convolution 스케줄링 기법
이영현(Younghyun Lee),김혜준(Hyejun Kim),유용승(Yongseung Yu),조명진(Myeongjin Cho),서지원(Jiwon Seo),박영준(Yongjun Park) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.11
As the AI industry evolves, neural network processing units (NPUs) are being developed to deliver AI services faster and more efficiently. One of the most important challenges for these NPUs is task scheduling to minimize off-chip memory accesses, which incur significant performance overhead. In particular, convolutional layers can be fused with multiple layers to reduce the memory accesses, but it is difficult to find the optimal schedule due to the too large exploration space. In this paper, we propose an efficient schedule exploration algorithm to optimize the fusion of multiple convolutional layers in NPUs. The proposed algorithm organizes the fusion group exploration space in the form of a grid to explore the optimal schedule. Experimental results show that the fusion schedule explored by the proposed method reduces the latency by 7.7% and reduces the off-chip memory access by 15% compared to the baseline algorithm.