SkyPilot argues that scheduling GB200 and GB300 NVL72 systems requires awareness of intra-rack connections and failure reserves, not just GPU counts. In a cited model with 2% of nodes unavailable across 128 racks, using 16 nodes plus two spares per rack instead of requiring all 18 increases usable nodes from 1,602 to 2,037, about 27%. That is not a speedup of individual GPUs. In a separate GR00T N1.7 experiment on 512 GB300 GPUs, rack-aware placement improved throughput by 14.4% over arbitrary placement. NVIDIA’s Slurm guidance likewise emphasizes rack boundaries and topology constraints. Operators need to consider model size, communication paths and whether jobs can safely restart.
Back to Tech
Blackwell scheduling must account for rack topology and spare capacity
A modeled 27% gain in usable capacity is distinct from a measured 14.4% training-throughput improvement.