[porker]

Good. Hopefully this will reduce the randomness of type-of-GPU assignment on the Pro plan.

I fine-tuned a model on Colab Pro earlier this year and having to launch and quit 6 or 7 times to get a faster graphics card to ensure it completed within the time limit sucked.

Hope this will give more transparency into whether you are assigned a whole card or a virtual slice of one. Something I could never work out before!

[cperry]

You're always getting the whole card, never a slice fwiw. We haven't found a GPU virtualization solution that has a strong enough security boundary between slices, so we keep you isolated.

And yes, hoping to give you more control over chip type too, stay tuned.

[porker]

Thanks that's good to know.

I would have upgraded to Pro+ if I had confidence it would speed up the process, but the promises of what you get (beyond "runs in background") were/are so vague I couldn't tell if it was worth it. I was only using 50-80 hours a month and it sounded like a plan aimed at more usage.

Google was/is leaving money on the table with the old scheme.