- GPU memory. Estimate the memory your model needs using the precision-based formulas (FP16, INT8, INT4) and the training overhead multiplier documented in Selecting an instance, then pick an instance whose aggregate GPU memory exceeds that.
- Regional availability. Not every instance is in every Availability Zone. Check the Instance Availability Matrix to confirm your target instance is in your target region.
- Capacity plan. Reserved Instances, Flex Reservations, On-Demand, and Spot trade cost for guarantees in different ways. See Capacity plans.
Administrator