Hey everyone! 👋
As a ML/Dev Ops engineer, everyday increasingly I am seeing a large number power hungry workloads, different models being trained at the same time. That coupled the current issues of climate change: it got me thinking about the impact of these types of workloads on the environment.
For example, I have seen CV models that have to be trained slowly for 2+ weeks on multiple GPUs and rinse and repeat during a full year.
I was wondering if anyone else feels like the tradeoff of model accuracy/performance against impact on the environment is not being talked about too much.
I'd love to validate the value of knowing and offsetting one's cloud CO2 footprint.
Thanks for your time!
Interesting topic 🤔
I use a lot of ML models / techniques in my projects and have also thought about the impact model training has on the environment when I first dipped my toes into Deep Learning via Fast.ai.
To me it felt like a waste of time and energy to train a model which can tell different puppy breeds apart on GPUs in cloud datacenters. That's even more true when a lot of students do that just to follow the course instructor.
I think it's an important topic and I'd love to hear your thoughts as to how you'd monetize it (if at all). Are you thinking about a non-profit initiative? What would be your overall goals?
I could imagine that it could be (unfortunately) a harder sell to companies given that it's mostly not a top 3 priority and oftentimes only used to get positive press.
Thanks for your well thought out response.
I truly think this is a topic that will gain traction in the next year, your point about doing the same training over and over is interesting. Wonder if there is a way to avoid that but somehow get you the learnings in a course like fast.ai?
I am thinking of building such a tool and give it away to generate exposure to this and somehow tie back cost analysis as costs follow carbon footprints pretty closely.
Yes, good question. When I took the course the students were encouraged to recreate the lessons from scratch after watching them (without referencing the material). This is a great way of learning but as you pointed out might be bad for the environment.
An interesting advancement in ML is transfer learning where one can use and build upon an existing model. This might help in reducing the carbon footprint in the future.