Google Collab

Google Collab is a website that allows you to edit and run Jupyter notebooks.

Google Collab is available in 3 tiers, including a free tier.

During testing, I found that Google Collab free-tier machines were slower than a regular MacBook Pro and less convenient to use as the editor is not as good as VSCode, but they are neat for sharing code easily and quickly.

Speed comparison

The first paid tier allows to use machines with very powerful GPUs including Nvidia A100.

Running inference tasks on this machine is about 16 times faster than a MacBook Pro The task was to complete a piece of code with 200 tokens. It took 135 seconds for the MacBook. It took 8.57 seconds for the A100.

Running training tasks on A100 is about 100 times faster. The task was to perform 400 training steps of the code generation model. It took about a week to run on the MacBook. It took 20 minutes for the A100.

Privacy issues and purpose

While Google Collab offers decent value for its cost (11€ / month), we cannot use it to handle sensitive data. Moreover, it is designed for experimentation and prototyping and cannot run for extended periods of time. However, Collab is very good for testing different parameters to training a neural network as training is very long otherwise. Once the best set of training parameters (hyperparameters) has been found, one can run the training on the MacBook, the slow way using real sensitive data.

Also, if needed, one can download the model trained on Collab to run inference on the Mac.