Hi all,
Just looking for advice on setting up Google Cloud machine for Colab - never used this before. Does it matter strongly which region I am in? As far as I can tell there are no GPUs available in the Australian regions. In US West I can get a Tesla P100, T4, or V100 - T4 being cheapest, can I get away with this?
Sorry for noob question, haven’t used Google Cloud before.
Cheers,
Ben
Hi Ben!
Don’t worry about asking these kinds of questions! It’s a confusing world to first walk into.
First of all Google Cloud (GCP) and Colab are separate products. Colab you can access any time at https://colab.research.google.com/ This is a simple Jupyter Notebook interface that you can access over the web. There are options to subscribe to various tiers of GPU “power” to make what you build run faster. Everything in Colab is essentially managed for you; you don’t have to install drivers or do (much) configuration other than select the “shape” of your infrastructure and then start running code. You do not have the ability to persistently store data (outside of Google Drive), nor can you choose memory, GPU, CPU type etc etc beyond some bland, nondescript categories.
In contradistinction, GCP is Google’s whole web services / infrastructure business. This is where, amongst many other services, you can build very specific machines with very specific configs to tailor to a particular workload. That service is known as Google Compute Engine (GCE). You “build” your machine and then remotely access it to run whatever jobs you have in mind. This means you also have to take care of installing drivers for the hardware you’ve configured and then also install all the software yourself. But it is customisable to whatever degree you desire.
A step up from this very granular control is GCP’s “Vertex AI” platform where, amongst other things, you can setup a web-based, pre-configured Jupyter Notebook. This will set up a Jupyter Notebook interface that is running on a hardware configuration you’ve chosen. It’s getting closer to Colab but you will still manage all your own storage, software, data etc etc.
If you want to persist with GCP, your best regions for GPUs close to Australia are Singapore and Japan where you can get multiple very powerful A100 GPUs if you wish. There are a few other GPU types that you can choose from – see here
As you’re just getting started, I would recommend you just work with Colab. It’s cheaper, fewer levels of technical depth to wade through and it “just works”.
If you’re interested in learning to build up and manage your own servers etc in the context of Fast AI, I can do no better than refer you to @jeremy 's live coding and tutorial sessions from Part 1 where he walks you through setting up machines on another Cloud provider (Paperspace Gradient). Those tutorials are super useful and generalisable to pretty much every other cloud platform.
Enjoy – if you get stuck, just post questions. For heavy duty, broad spectrum ML cloud environments, GCP is my preferred environment so, please feel free to reach out.
Hi Nick,
Thanks for the advice! I think I had a misunderstanding - I had some Google Cloud academic credits I wanted to put towards this and a couple of other projects, and I wanted to access these on Colab with a GCP VM. I didn’t realize that GCP credits can’t otherwise be applied to Colab subscriptions.
Thanks for the links. I work in physics and have previously only used university-run HPC servers and want to make a transition to using cloud providers, but am not really familiar with best practices in this domain. One consideration is that for my physics research in optics, it’s important to have good double-precision performance, so trying to figure out how best to access A100 or similar - T4 won’t cut it.
Cheers,
Ben
I’m not sure if it compares to the academic credits you have for GCP but lambda.ai has offered $150 credit to those participating in this part 2 of the course (there is a thread for it in part 2 forums.) You can play with A100 (or better) cards on their platform and I’m sure the lessons you learn would be easily transferable to other cloud providers if you choose to give it a shot.
All the best.