Spot vs Reserved GPUs: Duty Cycle and Checkpointing Decide the Cheaper Option

Teams often assume spot GPU instances are a simple win because of their steep discounts, but interruptions can erase those savings through lost work and restarts. A modeled comparison suggests spot is cheaper per delivered GPU-hour only when utilisation falls below roughly 65%, with reserved capacity winning above that level. The author argues most teams never measure their actual duty cycle before deciding. Checkpoint interval is identified as the most important adjustable parameter, with a minimum combined cost near a two-minute interval under the modeled assumptions. The piece also notes that workloads such as long non-resumable jobs, tightly coupled multi-node training, and latency-sensitive serving are poor fits for spot, while batch inference over fixed corpora suits it well.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.


Discussion (0)
Log in to join the discussion and vote.
Log in