Nodes

A node is a whole physical server assigned to one project. There is no hypervisor between your operating system and the hardware, and no other customer has a process, a virtual machine or a port on the same host.

What you get

Component	Detail
Accelerators	1, 2, 4 or 8 GPUs of a single model, depending on the plan
Host	Dedicated CPUs, memory and local NVMe listed on the plan
Network	One or more fabric ports per GPU on cluster plans, plus a management interface
Access	Root over SSH with keys you supply; serial console from the console or API
Image	Ubuntu 24.04 with current NVIDIA or ROCm drivers, or your own signed image

Plan identifiers follow cl.<accelerator>.x<count>, for example cl.b200.x8. The full list is in the GPU catalog.

Lifecycle

State	Meaning
provisioning	Drives are erased, firmware is re-flashed from signed images, the image is written
measuring	The node boots and records measurements of firmware, bootloader and kernel
ready	Evidence is available and the node accepts SSH
stopping	The OS is shut down; the node stays assigned to you
releasing	Drives are cryptographically erased; the record is kept in the node history

A stopped on-demand node is still billed, because the hardware remains reserved for you. Delete it to stop billing.

Images

You can boot the default image or register your own. Custom images must be signed with a key registered to the project; the signature is part of the boot measurements, so a policy can require it.

cloister images register ./ubuntu-cuda-12.8.qcow2 --sign-key ./image.key

Operator access

Cloister engineers do not have standing access to a node assigned to you. If a hardware fault needs hands-on work, an access request is opened against your project. Nothing happens until someone in your project approves it, and the session is logged to your audit trail.

Hardware faults

If a GPU or host component fails, we move you to a replacement node of the same plan. Local NVMe contents do not move with you — keep anything you need on block or object storage.