GPU Cloud

Define the workload before renting compute

Dedicated GPU compute gives you an environment to run your own software. Determine the exact model or dataset, framework version, peak device memory, CPU/RAM, disk requirements and access method first. GPU count means devices within one machine; it is not a distributed cluster or proof of a particular interconnect. Use the memory worksheet and Marketplace selection to prepare a shortlist.

Review an exact offering and authorization

Compare current GPU variant, count, memory, region, capacity class, CPU, RAM, storage and compatible environment. The public directory groups comparison rows; API deployment requires a current authenticated offering ID. A displayed rate and future availability are not guaranteed by an earlier screenshot.

Review the selected customer rate separately from the maximum authorized rate. API automation uses an explicit quote intent before launching the reviewed configuration and ceiling; omitted intent defaults to launch. A changed ceiling requires a fresh review rather than silently increasing it. Read quote and launch fields. An accepted launch is not proof of readiness.

Verify the delivered environment

1. Wait for readiness and use the assigned host, user and port from the workspace.
2. Verify the server fingerprint through the trusted workspace before accepting the SSH connection. Upload only your public key.
3. Inspect the allocated devices, memory, framework and disk. Run a small application-level check before a full job; NVIDIA SMI alone does not verify CUDA arithmetic.
4. Keep development listeners on loopback and use SSH forwarding. Confirm storage behavior before relying on local checkpoints.

Use the connection checklist and a bounded tensor test. A root shell inside a container is not permission to replace the host driver. Escalate managed-host failures instead of applying blanket driver, firmware or Docker changes.

Operate and finish deliberately

Bound job duration, retries and checkpoint intervals. Record instance ID and UTC times so usage windows can be matched to work. Before termination, export needed outputs and verify that they can be read from their destination. Do not assume every disk persists after termination or that an idle process triggers automatic shutdown.

Use the supported instance termination control and confirm resulting status. Closing SSH or stopping a notebook does not terminate the allocation. Reservations differ from completed charges; reconcile usage and wallet records without counting both as separate expenses. See instance controls and billing.

When to use another path

If you only need a supported model request, compare serverless inference before taking on driver, runtime, service security and storage ownership. Dedicated compute is appropriate when your requirements justify that work, not automatically faster or cheaper for every task.