Skip to content- Sources: primary, discussion
- Summary: SkyPilot reports about 2% node unavailability on production GB200 NVL72 clusters and, using NVIDIA's capacity model, finds 16-node slices give 2,037 usable nodes against 1,602 for 18-node slices on a 128-rack cluster. Its scheduler keeps each slice on one NVLink rack, hot swaps failed nodes for in-rack spares, and preempts checkpointable jobs to defragment racks. NVIDIA's GR00T team measured 14.4% higher training throughput with rack-aware placement on 512 GB300 GPUs.
- Why it matters: Teams running NVL72 clusters should size jobs below a full rack and use topology-aware placement to keep tensor and expert parallel traffic off InfiniBand.