Selecting and Verifying a GPU Runtime in Google Colab for Deep Learning

Problem: Slow CPU-only training in Colab

When you open a new notebook the default runtime runs on CPU, which can make even modest deep-learning workloads take minutes or hours.

Thesis: Pick a GPU runtime, verify it works, and guard against session limits

1. Selecting the GPU runtime

From the menu choose Runtime → Change runtime type. In the dialog set Hardware accelerator to GPU and click Save. Colab will provision a shared GPU, usually a Tesla T4 or P100, for the notebook.

2. Verifying that the GPU is visible

Run the following command in a code cell:

!nvidia-smi

The output should list a GPU name, driver version, and memory usage. If you see a line like Tesla T4 with non-zero memory, the accelerator is attached.

You can also ask TensorFlow or PyTorch directly:

import tensorflow as tf
print(tf.config.list_physical_devices('GPU'))

or

import torch
print(torch.cuda.is_available(), torch.cuda.get_device_name(0))

These calls should return a non-empty list or True and a device name.

3. Managing session limits and persisting work

Colab recycles the runtime after about 90 minutes of inactivity or after a maximum of 12 hours of continuous use. When that happens the GPU is released and any in-memory state is lost.

To avoid losing progress:

• Periodically save model checkpoints to Google Drive.
• Mount your Drive once per session:

# Create a mount point, run once
!mkdir -p drive
# Mount Drive, you will be prompted to authorize
!google-drive-ocamlfuse drive

After mounting, you can copy files:

# Example: save a PyTorch checkpoint
import torch
torch.save(model.state_dict(), 'drive/checkpoints/model_epoch_5.pt')

When the session timer in the top-right bar approaches the 90-minute idle warning, execute the copy step and then reload the notebook if needed.

Trade-off and limitation

Free Colab GPUs are shared resources. During peak demand the runtime selector may show a warning and fall back to CPU, or the selected GPU may have less memory than expected, about 15 GB. Models that exceed this limit will trigger out-of-memory errors; you must reduce batch size, use gradient checkpointing, or switch to a paid tier.

Actionable closing

Before starting a deep-learning experiment:

1. Open Runtime → Change runtime type and set GPU.
2. Run !nvidia-smi to confirm the accelerator.
3. Run a quick TensorFlow/PyTorch device check.
4. Mount Google Drive with the two-line snippet above.
5. Set up a checkpoint callback that writes to drive/checkpoints/ every N epochs or minutes.
6. Watch the session timer; when you see the idle warning, save your work and be ready for a possible recycle.

Following these steps lets you reap the speed-up of a GPU while protecting your work against Colab’s automatic recycling.