Is there a Runtime limit on pro/growth plan?

Probably a question asked about a million times.
When buying a subscription. Like the “pro” plan or the “growth” plan. Will there be a runtime/session limit on the gpu?

Im coming from Kaggle where the sessions last between 6 to 9 hours and then just stops. Thats a problem, because then I have to break up my machine learning projects. Which runs a lot longer than that.

So, I need it to not be a session limit. Is that the case here?

Say I buy the “pro” plan. Can I then use a “free” gpu as long as I need without it stopping and making my life really hard?

Hi @MyhreS All free instance types have a 6-hour runtime limit. You could potentially use the free type to prototype/develop your model and then switch to a paid instance type to run your long-running training. Paid instances do not have any runtime limits.

Ye, that was what I excpected. I run neural architecture search that goes over multiple days. So then I would have to split it up to take like 3 models at the time. Oh well, guess the service sadly doesn’t fit my needs.
Thanks for the answer!

Are paid instances not an option for you? Some of them e.g. the new A100 could potentially cut your training time down by a substantial amount.

It would be really nice. But lets say the gpu costs around 1 dollar pr hour. Then I need it for 100 hours. It would be 100 dollars.

Currently I run a (or multiple) simple neural architecture search because of the limit on kaggle. So I can only run like 20 models. That I think is not that many. So if I got access to a better GPU I would just train more complexe searches (that is what im looking for). So I would be using the same amount of time with a better gpu.

What im trying to say that I would need like 100 hours a month. That is 100 bucks. And im a student and doing this for fun. I would be willing to pay like 50 dollars each month in a subscription. Where I then wont be afraid to train models / do neural architecture searches because it wont be more expencive than that.

So, that is my reason. Did it make any sense?

But, could I ask you some other questions?

  1. When the session stops (using free gpu) will it delete datasets/outputs and that stuff?
  2. Is it possible to upload / use datasets of like some gb in the gradient notebook?
  3. To get access to “free” A100, what subscription-plan must be bought? (Is there a list that shows what “free” gpu’s that are usable at each plan?
  1. No, your files are saved in persistent storage and will be accessible from a subsequent run. Additionally, you can share files across other notebooks by storing data in the /storage directory. There’s more info here.
  2. You can upload data using the upload button in the notebook interface.
  3. There’s a list here. FYI we are just rolling out the Ampere generation of GPUs and have very limited supply on the free pool which uses our idle capacity. It may be a bit before the free Ampere GPUs are readily available. Other types (both paid and free) should be readily available.

Ok ok. Thats nice to know. Thanks. One last question. Say that I run a model to completion, and I have no further tasks to be done. Will the session stop automatically? So the next time I hit “run” it will start a new session and I’ll have 6 new hours

The session will stop if the notebook is fully idle (you are not coding and the instance isn’t executing any cells/code) or if you instruct it to stop after training is complete via some trigger in your code. Other than that, we won’t stop the instance until the auto-shutdown limit is met. Assuming it does stop, you can just hit start again to begin a new session. There are no limits to the overall number of sessions/hours you use – the 6-hour limit is just a constraint on the individual session.

Ok ok. Thanks, I’ll see if I can get everything running. Thanks for the answers. Been very helpfull :slight_smile: