//
Busque em todas as páginas da documentação
7 pages in this section.
Cost/performance trade-offs across accelerator families via SDK - when to provision GPU instances versus Trainium or Inferentia, and how the choice is just an EC2 instance type.
Launch a single P5/P6 or Trainium/Inferentia instance via SDK: check instance type availability, launch, describe, tag, and wait for readiness in both SDKs.
Reserving scarce accelerator capacity ahead of training runs: On-Demand Capacity Reservations versus Capacity Blocks for ML, via CreateCapacityReservation and PurchaseCapacityBlock.
Capacity planning and cost controls for accelerator-heavy workloads: quota requests, capacity reservations, tagging, EFA networking, cleanup, and Neuron validation habits via SDK.