Colab vs Paperspace Gradient
Which of these low-cost services for ML experiments is better?
I have been using Google Colab for my machine learning experiments for a very long time. Recently, when Colab turned out to be insufficient, I was forced to look for an alternative and I picked Paperspace Gradient.
How are they different in terms of costs, (GPU) power and other features?
Google Colab
I wrote about using Colab for my projects in my previous post. And I still consider it to be one of the best tools for both machine learning beginners as well as experts. Everything seems so trivial there: set up the environment, upload data (e.g. from Google Drive), write and run your code, share it with others (not only to read but also to collaborate).
💵 How much does it cost?
Google Colab is totally free. You don’t have to pay for running experiments on their GPU and your code can run for at most 12 hours, then the session will be terminated.
Unless you decided to use Colab Pro which costs $9,99/month and:
- gives you longer runtime (24 hours instead of 12),
- gives you better GPUs (e.g. T4/P100 instead of K80s),
- gives you 2x the memory of standard Colab VMs and 2x as many CPUs,
- is available only in US and Canada (as for right now).
🚀 How much power do I get?
Well, it depends. Colab FAQ states that you can get various types of GPU (GPUs available in Colab often include Nvidia K80s, T4s, P4s and P100s). It is never guaranteed which one do you get… and for how long.
What does that mean? Colab is well-known for its “dynamic usage limits” and this can be really confusing for some people, so let me explain. Colab gives you their GPU resources for free, you can use it for 12 hours, but…
- there’s no way to choose which GPU you will connect to,
- you will be disconnected after idle time (90 mins but it may vary),
- in the middle of session, you may be told that GPU is unavailable,
These things but also “other factors” vary over time — says Colab FAQ.
🎁 Other features
Google Colab is all about running code in Jupyter notebooks. You literally upload data (from Google Drive or directly in Colab), run code cells and then eventually share your notebook with others (or just download it).
Other features? None that I can think of, but I believe it is also an advantage if you are just looking for a simple tool to run your Jupyter code and you don’t need CI, experiment tracking, project management and so on — unlike in the case of Paperspace.
Paperspace Gradient
I have to admit that I am relatively new to Paperspace, so I might have not discovered all of its features yet. But after few days or weeks of using it as the alternative for Colab, I think I can already tell the diference(s).
Firstly, let me describe Gradient in the same categories as before (costs and “power”). Then I’ll move on to other functionalities available there and this time, the list is going to be really long.
💵 How much does it cost?
Paperspace Gradient comes with three pricing plans (plus one for Enterprise), one of which is free. In the Free plan you are allowed to:
- store up to 5GB in dedicated persistent storage,
- use free tier CPU/GPU i.e. C3 CPU, NVIDIA M4000 and P5000 GPUs,
- run your code for at most 6 hours when using free CPU/GPU instances,
- use only public notebooks (which cannot be set private in Free tier).
That sounds really neat. If you decide to upgrade your subscription to either G1 developer ($8/month) or G2 professional plan ($24/month), you can get:
- private notebooks,
- better CPUs and GPUs depending on your tier, with a full list here,
- 200GB or even 1TB of persistent storage instead from 5GB in Free tier,
- bigger limits on number of notebooks (running and total) and jobs.
🚀 How much power do I get?
There are plenty of CPU and GPU instances to choose from (and you can literally choose — unlike in Colab, where you never know what type of GPU you’ll get). But there is one important thing.
Although more instances are available in higher subscription plans, it does not mean they are free. There are only three instances that you can use at no cost (C3, M4000, P5000). In the table above, ticks under pricing plans mean that these GPUs are available for you, but you still need to pay to use them (usually few cents per hour).
Moreover than that, if you subscribe to e.g. G1 plan but you decide to use one of Free instances, your notebook cannot be set to private and it will have a 6-hour limit (auto-shutdown). It is just because of using free CPUs/GPUs. These restrictions do not apply to other instances available in your tier.
🎁 Other features
What is the essence of Colab (just running Jupyter notebooks) is not even a tenth of what Paperspace Gradient has to offer. So starting with notebooks — you can run them on CPU, GPU or TPU using pre-built notebook containers as well as your custom ones (or just start with clean environment).
Authors of Gradient also provide ML Showcase — a curated list of interactive machine learning projects build on top of Gradient notebooks. You can explore these notebooks and play with them.
But let’s go back to built-in features. In Gradient Docs we read about:
- Experiments that can be run both on a single-node and in a distributed environments. According to docs, Gradient experiments support HPO (Hyperparameter Optimization) and metrics logging. I haven’t tested it yet, but that sounds really interesting to have all of that in one place.
- Models repository where you will be able store artifacts, along with their summary (loss, accuracy) and tags. You can read how to export TF model that can be stored in Gradient hub and then used for deployment.
- Deployment option for trained models and Projects that serve as a workspace for storing your experiments, artifacts and deployed models together in one place.
and other functionalities (some of which only available in private clusters).
Conclusion
- Paperspace Gradient offers much more than Google Colab does, if you want to manage your projects and models instead of just running computations.
- Theoretically you can get better GPUs for free in Colab (Colab FAQ mentions K80 or P100s that are available in Mid or High tiers of Gradient), but on the other hand — you are never guaranteed which ones you get.
- In Colab, you are not even guaranteed that CPU/TPU instance will be given to you for the whole session (i.e. 12 hours) and you will be kicked out if being idle for 90 minutes — more or less, because so many factors vary…
- Gradient offers a persistent storage. It means that you don’t need to upload data everytime you run notebook. This can save you a lot of time when processing large datasets. However, you are only given 5GB of free space. Upgrading it to 200GB will cost you $8 per month.
I will keep using Google Colab on daily basis to perform light computations (when I don’t mind if it gets disconnected after a couple of hours). It comes with a bigger storage, Google Drive integration and (in my opinion) much simpler UI.
When that’s enough and I need 6-hour long session that cannot be interrupted, Paperspace is the answer for me. At least as long as I don’t care whether my notebooks is public or not…
