With GPUs, the agent doesn't run copies of the same loop. Instead, it submits batches of experiments in parallel. Each experiment still runs on a single GPU for minutes. But now experiments run at the same time.
The throughput jump is immediate. A single GPU gives you about experiments per hour. With GPUs, the SkyPilot run hit roughly experiments per hour. That's a x speedup in wall-clock time. The same search that takes hours sequentially finishes in about .