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

Answer

For idle periods shorter than the typical ~90‑minute free‑tier timeout, the GPU remains allocated to your notebook, so kernel execution continues on the GPU. However, after a period of inactivity Colab may throttle or reclaim the GPU, which produces a measurable slowdown when the notebook resumes.

Confirmed facts

• Google Colab’s free tier provides a Tesla T4 or K80 GPU with usage limits and may reclaim or throttle it after extended idle periods to serve other users.
• When the GPU is reclaimed, subsequent code execution incurs re‑initialization overhead, resulting in a measurable performance drop after the notebook resumes from idle.

Likely explanation

The underlying VM enters a low‑power or suspended state after inactivity; upon resume the GPU context is reset, causing slower kernel launch times until caches warm up. Colab may also pre‑empt the free‑tier GPU for other users, moving the notebook to a less performant GPU or temporarily falling back to CPU execution.

Steps to verify and mitigate

1. Before an idle period, run !nvidia-smi to note the GPU name, memory usage and utilization.
2. Run a small, repeatable GPU benchmark (e.g., multiplying two 1024×1024 random matrices with TensorFlow/PyTorch) and record its execution time.
3. Leave the notebook idle for the duration you wish to test.
4. After resuming, run !nvidia-smi again and repeat the benchmark.
5. If the benchmark time has increased significantly or nvidia-smi shows a different GPU or low utilization, the GPU was likely throttled or reclaimed.
6. To keep the GPU active, execute a lightweight GPU operation (such as !nvidia-smi or a tiny tensor multiply) every few minutes, or manually reconnect the runtime via the “Runtime → Reconnect” menu if performance degrades.

Missing diagnostic detail

To refine the recommendation, please specify the idle length at which you first observed the slowdown (e.g., 15 min, 30 min, 60 min). This helps determine whether the issue stems from throttling or full reclamation.