Deep learning has a terrible carbon footprint
One of the pressing problems to be solved until 2030, is the climate change and the transformation of current technologies towards zero emission. This must be seen in the light that among others the EU member states have signed the 2015 Paris Climate Agreement. The question now is whether computer science (CS) is part of the solution or part of the problem. It turns out that CS, especially the growth in AI applications will tremendously increase the demand for supercomputing capacity. This, in combination with blockchain technology, has the tendency that energy consumption by digital technology will go through the roof. There are predictions that by 2040 the energy required for computation will have exceeded the world energy production unless we find energy saving technologies that add up to the problem.
A recent article in the MIT Technology Review /1/ reports about the carbon print caused by artificial intelligence shows the current limitation of this technology.
The artificial-intelligence industry is often compared to the oil industry: once mined and refined, data, like oil, can be a highly lucrative commodity. Now it seems the metaphor may extend even further. Like its fossil-fuel counterpart, the process of deep learning has an outsize environmental impact.
In a new paper, /2/ researchers at the University of Massachusetts, Amherst, performed a life cycle assessment for training several common large AI models. They found that the process can emit more than 626,000 pounds of carbon dioxide equivalent—nearly five times the lifetime emissions of the average American car (and that includes manufacture of the car itself).
As a consequence
- Research is therefore needed in the field of chips and processors with at least an order of magnitude less energy consumption
- Research is needed in the field of high performance operating systems (may be special purpose) that deliver results faster at lower (energy-)costs.
- May be an interdisciplinary approach between CS and the semiconductor industry should be pursued to address the issues to minimize energy consumption.
- We have also to address the problem of waste. We need to develop technologies that deal with end-of-live processing of computing equipment. Both in terms of reuse of materials (such as precious metals) as well as avoidance of soil pollution with the leftovers.
MIT Technology Review:
Research Paper “Energy and Policy Considerations for Deep Learning in NLP”, by Emma Strubell, Ananya Ganesh, Andrew McCallum