Buy the fabric and scheduler promise.

Video caption

Buy the fabric and scheduler promise.

The requested topology may not be available.

Start with vendor terms; inspect traces in a pilot.

My rule: Compare the service the workload consumes.

#EricFieldNotes

Full written post / accessibility read

Neocloud procurement often collapses to GPU count and hourly price. For a distributed job, I would buy a placement class, start-time envelope, communication path, recovery behavior and cost per completed result. Those determine whether capacity is usable.

Imagine a provider whose fleet is mostly occupied by single-device inference. The remaining cards are scattered. Your training job needs a connected allocation by Monday. The provider's average utilization can be excellent while your job waits and the promised deadline fails.

Before a contract, ask for the placement class, service terms and sample traces the vendor can share. In a scoped pilot, retain admission and placement decisions, device and NIC topology, collective diagnostics, application step time, checkpoint behavior and bill. Request a contended run only where the pilot terms permit it.

Contract for topology-valid start, tail completion and recovery, then compare delivered cost. Do this because a cheap device hour that cannot be assembled into your job can be more expensive than a dearer, predictable allocation.

#EricFieldNotes

Four-beat scene transcript

1. Buy the fabric and scheduler promise.

Neocloud procurement often collapses to GPU count and hourly price. For a distributed job, I would buy a placement class, start-time envelope, communication path, recovery behavior and cost per completed result. Those determine whether capacity is usable.

Visual: A utilization headline cannot guarantee a run.

2. A fleet can be busy and still miss you.

Imagine a provider whose fleet is mostly occupied by single-device inference. The remaining cards are scattered. Your training job needs a connected allocation by Monday. The provider's average utilization can be excellent while your job waits and the promised deadline fails.

Visual: The requested topology may not be available.

3. Stage an evidence packet.

Before a contract, ask for the placement class, service terms and sample traces the vendor can share. In a scoped pilot, retain admission and placement decisions, device and NIC topology, collective diagnostics, application step time, checkpoint behavior and bill. Request a contended run only where the pilot terms permit it.

Visual: Start with vendor terms; inspect traces in a pilot.

4. Price connected capacity.

Contract for topology-valid start, tail completion and recovery, then compare delivered cost. Do this because a cheap device hour that cannot be assembled into your job can be more expensive than a dearer, predictable allocation.

Visual: Compare the service the workload consumes.

Research and claim limits

• NVIDIA NCCL performance and tuning (S122)
• Kueue overview (S124)
• NVIDIA Fabric Manager (S126)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.