FP8–FP16 혼합 정밀도 기반의 MobileNetV1 추론 성능 분석|RISS 상세보기

FP8–FP16 혼합 정밀도 기반의 MobileNetV1 추론 성능 분석

한글로보기

https://www.riss.kr/link?id=A109889330

저자

김재성(Jaeseong Kim) ; 이채빈(Chaebin Lee) ; 이다영(Dayoung Lee) ; 이승은(Seung Eun Lee)
발행기관
대한전자공학회
학술지명
대한전자공학회 학술대회
권호사항

Vol.2025 No.6 [2025]
발행연도
2025
작성언어
Korean
KDC
569
자료형태
학술저널
수록면

2189-2192(4쪽)
제공처
DBpia

0
상세조회
0
다운로드
0
내보내기

부가정보

다국어 초록 (Multilingual Abstract)

To enable high-performance AI inference in resource-constrained environments such as embedded systems, mixed-precision inference techniques have been actively studied to reduce memory usage. In particular, the combination of FP8 and FP16 is being explored as a means to maximize memory efficiency. However, the limited dynamic range of FP8 can lead to significant accuracy degradation. In this paper, we apply mixed-precision inference using FP8-FP16 to the core operations of MobileNetV1—depthwise and pointwise convolutions—and analyze the trade-off between inference accuracy and memory usage based on the scope of precision reduction.

번역하기

To enable high-performance AI inference in resource-constrained environments such as embedded systems, mixed-precision inference techniques have been actively studied to reduce memory usage. In particular, the combination of FP8 and FP16 is being expl...

To enable high-performance AI inference in resource-constrained environments such as embedded systems, mixed-precision inference techniques have been actively studied to reduce memory usage. In particular, the combination of FP8 and FP16 is being explored as a means to maximize memory efficiency. However, the limited dynamic range of FP8 can lead to significant accuracy degradation. In this paper, we apply mixed-precision inference using FP8-FP16 to the core operations of MobileNetV1—depthwise and pointwise convolutions—and analyze the trade-off between inference accuracy and memory usage based on the scope of precision reduction.