Its model favors 16-node slices with two spare nodes per rack; a separate GR00T training test reported 14.4% higher throughput.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
Primary source: [Hacker News](https://skypilot.ai/blog/grace-blackwell-needs-a-new-scheduler)
Why it matters #
NVL72 turns a rack into a high-bandwidth computing domain, so scheduler decisions can affect both job speed and how much capacity survives failures.
SkyPilot says its scheduler can make 27% more capacity usable in NVIDIA NVL72 clusters by leaving two nodes in each rack available as spares. The claim, published in a technical post on October 29th, is about how operators allocate existing machines, not a hardware upgrade: in the model SkyPilot cites, 16-node job slices yield 2,037 usable nodes across 128 racks, compared with 1,602 when jobs consume all 18 nodes in each rack.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.