MOST: Memory Oversubscription-Aware Scheduling for Tensor Migration on GPU Unified Storage
Sign Up to like & getrecommendations! Published in 2025 at "IEEE Computer Architecture Letters"
DOI: 10.1109/lca.2025.3580264
Abstract: Deep Neural Network (DNN) training demands large memory capacities that exceed the limits of current GPU onboard memory. Expanding GPU memory with SSDs is a cost-effective approach. However, the low bandwidth of SSDs introduces severe… read more here.
Keywords: memory; migration; oversubscription aware; memory oversubscription ... See more keywords