
SkyPilot says rack-local scheduling yields 27% more usable GPUs
Topology-aware placement, in-place recovery, and defragmentation keep multi-node jobs on NVL72 NVLink instead of InfiniBand.
3 stories, newest first · one card per story

Topology-aware placement, in-place recovery, and defragmentation keep multi-node jobs on NVL72 NVLink instead of InfiniBand.
Across 12 reported settings, TAO cuts latency by up to 26.8% and mean TTFT by up to 36.3% versus the best baseline for each metric.
PEARL adapts prefill-decode execution and elastic GPUs, raising agentic RL throughput by up to 2.79x.
Today 2.6K readers · 2.8K views · 145 reads · all time 206K views