Source: Deep Learning on Medium
Artificial Intelligence: a boon or a bane for our planet?
“Global warming is a total, and very expensive, hoax! “– Donald Trump
If you live on this planet, you already know how untrue the statement is.
Global warming is a real, dangerous and man-made phenomenon. With its aggravating effects on the climate of the earth, we will soon run out of time to unroll it or at the very least, slow it down. Fortunately, Artificial Intelligence (AI) is here to address the problem. With its promising results in the fields of agriculture, renewable energy and wildlife protection, companies like Microsoft, Google are investing millions of dollars to research new applications for AI over the next few decades.
But success comes with a price. A very recent research on energy consumption of AI models, has revealed their substantial negative impacts on our environment. In addition to the cost of hardware and electricity, and compute time, the carbon footprint required to fuel modern tensor processing hardware is considerable. Naturally, the questions arise:
1. With respect to the above findings, can AI save our climate?
2. If yes, how so?
Environmental Footprint of AI
Since last three decades, the advancement of AI has been incredible. And lost in amazement of our inventions, we never noticed that carbon footprints can be invisible(in literal sense). We don’t see spewing black smoke in the air or the early-morning smog, caused by the information and technology centers, but they are quite guilty of not being environment friendly.
You need scooters to ride through the corridors of Data Centers!
Upload a photo in Instagram and the possibility is, the photo gets stored in one of the data centers in EU, because apparently, its US data centers have run out of space. Stacks of millions of circuit-boards in these data centers are machines, loaded with petabytes of data and doing computations. With growing demand, this will lead to an explosion in energy use.
An Energy Forecast Projection by Swedish researcher Anders Andrae suggests that by 2030 electricity use by information and communication technology (ICT) could exceed 20% of the global total, with data centers using more than one-third of that. Currently, the data centers collectively use an estimated 200-terawatt hours (TWh) each year. That is more than the total energy consumption of some countries, including Iran.
General neural network architectures
Deep learning , being the buzz word of AI, every other science minded person is trying their hands at it. From my personal experience, we read and implement a paper or fork a model architecture from github, to use it our way. With almost no detailed information on the models available, we try training and re-training with new data and hyperparameters, over and over. The carbon emission or energy consumption is never paid heed to.
For instance, consider training a standard neural net inception on imagenet to 95% accuracy of object detection. It takes around 40 GPU-hours consuming around 10 kWh and produce around 10 lbs of CO2, which is equivalent to around 2–3 hours of running a central air conditioner .
Now consider, 50000 (I am sure there’s much more) people trying similar actions individually. If not now, this is going to be a problem in the near future.
Neural networks in a bigger scale
According to a recent study by the University of Massachusetts, training a Transformer-based neural network architecture for building contextual representations, along with small hyperparameter grid searches, can emit more than 626,000 pounds of carbon dioxide equivalent. That’s about five times the lifetime emissions of an average American car, including its manufacture. The study also reports that the model emits substantial carbon emissions; training BERT on GPU is roughly equivalent to a trans-American flight.
Model inference and hyperparameter tuning are the culprits
Today a bigger source of energy consumption comes after the models are deployed. In deep learning applications, inference drives as much as 90% of the compute costs of the application. That is mostly because inference happens on a single input in real time, consuming only a small amount of GPU compute. So, even at peak load, GPU compute capacity is not used 100%, which is uneconomical.
Tuning is a part of training a model from scratch. With known hyperparameter values on a given data, this is not an issue. But with complex architectures, the cost of tuning with new data will be a problem to look up to in future.
Why is this important?
Evidently, AI is turning into a potential source of energy consumption and greenhouse gas emitter. With not much research being done on measuring the carbon and dollar cost of working with computation intensive AI models, it is hard to predict the exact number. Regardless, the numbers we have and the projected energy demand in 20 years are good concerns because:
- Energy is not currently derived from carbon-neutral sources in many locations
- Why not use the energy required to train/infer a model to light up a remote village instead?