Free‑tier GPU runtime shows performance drop after idle periods in Google Colab

Goal: Determine whether a Google Colab free‑tier GPU runtime continues to provide GPU acceleration after a period of inactivity, or whether the service silently falls back to a CPU‑only execution environment.

Constraints: The free tier’s inactivity timeout is not precisely specified and may vary with overall platform load; exceeding the free‑tier GPU quota can trigger a runtime switch without explicit notification, making it difficult to predict performance for low‑traffic, intermittent GPU jobs.

Questions: Does the GPU remain allocated after idle periods shorter than the typical 90‑minute timeout? Is there a measurable drop in execution time that correlates with idle duration? At what idle length does a fallback to CPU‑only execution become likely?