GoKawiilSkyPilot platform adds scheduler to keep AI jobs rack-local on NVIDIA NVL72 systems
SkyPilot has detailed a scheduling approach designed to maximize the benefit of NVIDIA's Grace Blackwell NVL72 racks, which link 72 GPUs via a unified NVLink fabric instead of relying on InfiniBand across nodes. The company says its platform preserves rack-locality for multi-node training jobs through topology-aware placement, in-place recovery from hardware failures, and automatic defragmentation of shared clusters.
GoKawiil's interpretation of the reporting above, not reported fact.
Running distributed training over InfiniBand instead of NVLink can cut performance by 43%, according to the post, so losing rack-locality due to failures or scattered free capacity could erode much of the NVL72 hardware's advantage. This suggests that as GPU clusters scale, scheduling software may become as important as the underlying interconnect hardware in determining real-world training throughput.
- NVIDIA's NVL72 racks connect 72 GPUs with NVLink, offering far more bandwidth than InfiniBand for multi-node jobs.
- Workload placement across racks, not just hardware, determines whether NVL72's performance benefits are realized.
- SkyPilot's scheduler aims to maintain rack-locality via topology-aware placement, failure recovery, and defragmentation.
Source: skypilot.ai, 2026-10-29
Published there as: “Grace Blackwell needs a new scheduler: 27% more usable GPUs from the same racks”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.