Compute & memory
Specify the model, precision, framework, memory footprint, training or inference pattern, desired quantity, and growth plan. GPU memory alone does not determine total workload capacity.
INFRASTRUCTURE PLANNING GUIDE
The details behind
a successful deployment.
A practical starting point for defining your requirements. Exact capabilities depend on the selected hardware and agreed service scope.
Fabric & topology
Ethernet, InfiniBand, and RoCE are distinct network design options. GPU-to-GPU NVLink within a supported system is different from the inter-node cluster fabric.
Storage & data
Define dataset size, checkpoint behavior, throughput needs, access patterns, local storage, and external storage integration. Share data governance requirements during planning.
Environment & provisioning
Identify operating systems, drivers, containers, orchestration needs, access controls, and management nodes. Confirm support scope and responsibilities in the agreement.
Power & cooling
Hardware density, system power, airflow, and any liquid-cooling requirements must be assessed with the selected system and deployment region.
Validation & acceptance
Agree on hardware diagnostics, burn-in, communication tests, applicable NCCL checks, network and storage testing, and documented acceptance criteria.
TURN REQUIREMENTS INTO A PLAN