Topograph gives GPU schedulers a map of the network

NVIDIA Topograph discovers how cluster resources connect and translates that information into Kubernetes labels or Slurm configuration. Schedulers can then favour GPUs with shorter, faster communication paths instead of spreading a tightly coupled job across distant network links. Providers discover the topology; output engines adapt it to the workload manager, with refreshes triggered by supported cluster events.

NVIDIA Technical Blog · Sep 22, 2026

Why it matters

GPU placement can waste compute when communication crosses congested links.

More context

Try the simulation utilities before deploying. Kubernetes setup requires version 1.27 or later, Helm and a supported provider. KAI Scheduler or Kueue can consume the labels for placing a group of related Pods together.

Keep in mind

Open source tooling, with support varying by provider and version. It reflects reported topology; native Pod affinity is a preference, not globally optimal group placement. No performance measurements were reproduced.

Original read