Clusters

A cluster is a reserved set of eight-GPU nodes joined by a dedicated high-bandwidth fabric. The fabric ports, and the switches they connect to, belong to your cluster for the length of the reservation.

Sizes and terms

Clusters start at two nodes (16 GPUs) and are reserved for one to twelve months. Larger allocations are planned with you before the term starts.

Size	GPUs	Typical use
2–8 nodes	16–64	Fine-tuning and continued pre-training
9–32 nodes	72–256	Pre-training mid-sized models
33–128 nodes	264–1,024	Large pre-training runs

Fabric

Each GPU has its own 400 Gb/s port (800 Gb/s on B300 plans). Ports with the same GPU index across nodes connect to the same leaf switch — a rail-optimised layout that keeps most collective traffic one hop away. See Networking for the topology.

Scheduling

Clusters are delivered as plain nodes so you can run your own scheduler. Two managed options are available:

• Slurm: a controller node and accounting database inside your boundary.
• Kubernetes: a control plane on dedicated hosts, with the NVIDIA GPU operator installed.

cloister clusters create \
  --plan cl.h100.x8 --nodes 16 \
  --term 6m --region eu-n \
  --scheduler slurm

Health checks

Before a cluster is handed over, every node passes a burn-in: GPU memory tests, NVLink bandwidth checks and an all-reduce test across the fabric. The results are attached to the cluster record so you can compare them with your own runs.