Grace Blackwell needs a new scheduler: 27% more usable GPUs from the same racks
89 points by wondersky
SkyPilot enhances GPU utilization by 27% through a topology-aware placement strategy that optimizes multi-node training and inference on NVLink. It achieves this by ensuring that only rack-local jobs benefit from speed improvements, employing in-place recovery for GPU failures, and implementing automatic defragmentation to maintain performance efficiency.