I have a basic laptop. I don't have a Cuda GPU or even a lot of memory, so that is limiting. I am trying a few different options right now.
I am wanting to try different things out, learn, do both training and inference.
Google Colab is good, but instances shut down, and you need to use the Jupyter-style interface whereas I prefer an IDE most of the time.
Modal.com is a cloud for running GPU workloads that is developer friendly is good, but you do need to learn their APIs. It can be fiddly sometimes to figure out how you do what you need to "the modal way".
I am happy to spend say up to $50/m on the hobby.
I will post updates sometimes on my thinking, and I won't bet.
1,000Runpod is a good one - very cheap, they also support Docker images. No free allowance though.
@osmarks, while cheapness is great there are a few things to consider:
I find for most things, I don't need an A100 40Gb anyway. T4 is plenty.
My spend on modal so far this week has been say $5 (they give $30/m free) and I haven't been "holding back". So free allowances make a big difference
Also the developer experience makes a big difference. At some point, even if someone offers me a free GPU, if it is a lot of hassle to use (unreliable, difficult to run workloads, etc.) then I probably wont use it.
A key consideration is volumes. If I can't persistently mount 20-200Gb of data then it just will be too painful to work with. If
First class container support is also big thing - Modal supports container images and it made a big difference. I need a CUDA driver? There is a container for that. It is less messing around. This is what made it much easier for me to get tinygrad running on modal with GPU support.
I would now put docker containers & volumes as must-haves.
Without docker containers it is too much dev work to do to get things in the correct state. Without volumes I pretty much need to keep the thing running 24/7 at which point I might as well buy a rig.
@Undox I am not saying Lambda doesn't offer those things - I still need to investigate. But I am just brain dumping to help traders decide.
@Undox They do have volumes. I don't know about the container support because I don't actually use cloud (my local hardware is good enough for most things I care about).
So what am I doing right now? I am using Modal. I was using Colab. Modal is a bit fiddly though - I am writing a script to sort of "chuck up the local code and run it" but there are some ops challenges, like dealing with 15Gb (maybe larger) weights files. But the only way to get rid of those challenges is a local machine, but that'll cost me $2000+, but maybe I will care enough by end of year to do that, but then I might like to use A100's every now and then. So really up in the air!