8GB 보급형 GPU 환경에서의 다단계 영상 처리 파이프라인 GPU 메모리 최적화

초록

Multi-stage image processing pipelines that chain multiple deep-learning models such as super-resolution, face enhancement, colorization, and image analysis implicitly assume high-end GPUs with 16GB or more of memory, limiting their use on commodity 8GB GPUs common among individual researchers and small institutions. This paper proposes a set of memory optimization techniques—stage-wise module unloading, dynamic tile sizing, exception-safe memory cleanup, and an input-size safety threshold—that enable such a four-stage pipeline to run on an 8GB GPU. Controlled experiments show that stage-wise unloading stabilizes single-image processing while keeping per-stage VRAM occupancy near a constant baseline. We further profile batch processing and attribute its out-of-memory failures to memory fragmentation rather than absolute capacity shortage. The proposed approach broadens the practical accessibility of such multi-stage pipelines to commodity hardware.

키워드