Introduction
A few notes on the new series of blogs I am starting, and on the contents of this post.
This is the first in a series of blogs I will be posting around the intersections of AI, Big Data, bioinformatics, and the environment. Why these? Two reasons. First and simplest, those four elements are at the center of my PhD work at Trinity College Dublin – over the next 4 years I will be delving very deeply into all of them.
Second: these are topics that, while of critical importance in the modern world, often have very high barriers to entry. The science, if even available to the general public (which is not always the case; much of our world’s best scientific work is hidden behind paywalls) is written to a specialist audience and not a general one.
This post will cover one aspect of AI and the Environment, focusing on the problem that exists. Later posts will go into solutions to this problem (there are many exciting ones) as well as reasons that even very electricity-hungry AI can do wonders for the environment. Here are the key take-aways:
- Complex AIs require loads of electricity, computer hardware, and time to create
- Making more complex AIs gives diminishing returns
- Most AI currently focuses only on results, not on environmental impact
Some Background
AIs are based on loads of math, and making computers do this math requires huge amounts of time and electricity.
AI, or Artificial Intelligence, is a set of methods to teach computers how to reason, act intelligently, or learn from data to do some job.
Goodness" and "effectiveness" of an AI can be measured in many ways; those details are outside the scope of this post. For now, you can think of every AI has having a job to do, and its "effectiveness" as being how well it does its job.
As to how it works: it all comes down to math. The building blocks of AI are mathematical operations – the details of that are out of the scope of this post, the most basic parts of AI (it’s "atoms" if you will) are multiplication, division, addition, and subtraction.
The amount of work needed to run AI is often measured in what are called "GPU-hours" – these are basically "man-hours", but for computers. The GPU (or Graphics Processing Unit) is the part of a computer that is able to perform math at incredible speeds, making AI possible.
Here’s the thing: every math operation requires electricity. Individually, they take practically nothing. But most of the best modern AIs operate on such huge amounts of operations that this adds up to a huge impact.
For example: Some of the best models (such as "RoBERTa28") can take on the order of 25,000 GPU-hours [1]. That’s around 2.85 GPU-years – meaning that it is equivalent to running a single top-of-the-line computer, at high power for almost 3 years, without ever turning it off.
But that’s not all. Sourcing computer parts (which require rare-earth metals such as gold) is an expensive task. These parts must be refined, transported, assembled, tested, and shipped to users. Then, the computers must be run with high electric power to create AI. Even once an AI is created, some require so much computational power to be used that they cannot be practically used on anything but the best computer hardware humanity has to offer [1].
As another example: the "AlphaGo" AI, trained to play the game Go (and far outdoes humans at this task) has an estimated cost of $35,000,000 to reproduce – one of the most expensive models to date [1].
"Red" AI
Red AI is AI that focusses only on results – not on resources consumed.
Schwartz at al. defined "Red AI" as AI the focused only on results – how effective the AI itself was – without taking environmental cost into account [1].
There are two major factors that make AIs "Red". The first is using larger datasets so the AI has more to learn from. The second is using more complex AI models (i.e., more math) so the AI learns more effectively.
These two approaches have two things in common. They both require huge amounts of resources. They both also suffer from diminishing returns.
Schwartz at al. explains that both of these approaches – more data and more computation – lead to only logarithmically better results. What does that mean?
Here’s a logarithm (or "log").

As x increases, it’s logarithm increases – but slower and slower over time. It’s a terrible case of diminishing returns. In our case "x" is the amount of data, or the amount of math, an AI has to do. Log(x) is how "good" the AI is.
This means that not only is Red AI resource-hungry, but it also gets more and more unsustainable the more you push for huge data and computational power.
Conclusion
By now, it should be clear that AI has some major problems in terms of resource consumption. Here are the key take-aways:
- Complex AIs require loads of electricity, computer hardware, and time to create
- Making more complex AIs gives diminishing returns
- Most AI currently focuses only on results, not on environmental impact
See you in the next post!
References
- Schwartz, Roy, et al. "Green ai." Communications of the ACM 63.12 (2020): 54–63.