Work out neural architecture search cost instantly with clear inputs, formula shown and shareable results.
Neural architecture search cost is candidates times training cost per candidate, and parallelism buys wall-clock time without reducing the bill. This is why proxy strategies dominate practice: training candidates for a few epochs, using weight sharing as in ENAS, or predicting final accuracy from early curves cuts hours per candidate by one to two orders of magnitude.
Search cost
GPU hours = candidates x hours per candidate; cost = GPU hours x hourly price; wall clock = GPU hours / parallel GPUs
Rarely for standard vision or language tasks, where published architectures are strong. It pays off when optimising for unusual hardware constraints such as a specific latency or memory budget.
Use a reduced proxy task, early stopping on poor candidates, weight sharing, and a search space constrained by prior knowledge instead of an open-ended one.