Grace Blackwell needs a new scheduler: 27% more usable GPUs from the same racks

NVIDIA's NVL72 systems connect 72 GPUs per rack with unified NVLink fabric, enabling significantly faster multi-node communication compared to traditional InfiniBand connections—with potential performance degradation of 43% when using InfiniBand instead. However, maximizing these benefits requires intelligent workload placement to keep jobs within a single rack, as multi-node jobs spanning multiple racks still face InfiniBand bottlenecks. SkyPilot Platform addresses this challenge through topology-aware scheduling, in-place failure recovery, and automatic cluster defragmentation to maintain rack-locality and improve usable GPU capacity by 27%.