Grace Blackwell needs a new scheduler: 27% more usable GPUs from the same racks
Summary
The article explains how SkyPilot leverages NVL72 rack-scale GPUs and topology-aware scheduling to increase usable GPUs from the same racks by about 27%. It discusses rack-local NVLink, the impact of fragmentation, and features like in-place recovery from GPU failures and automatic defragmentation to optimize large-scale AI training and inference workloads.