Researchers predict that by 2030, the industry could discard e-waste equivalent to more than 10 billion iPhones per year due to the enormous and rapidly evolving computing requirements of AI models. Masu.
In a paper published in the journal Nature, researchers from the University of Cambridge and the Chinese Academy of Sciences set out to predict how much e-waste this growing industry could generate. There is. Their aim is not to limit the adoption of this technology, emphasizing that it is initially promising and perhaps inevitable, but to move the world further toward tangible results from its rapid expansion. It is important to prepare well.
They explain that energy costs are being scrutinized because they are already being incurred.
However, less attention has been paid to the physical materials involved in their life cycles and the waste streams of used electronic devices.
The purpose of our study is not to accurately predict the volume of AI servers and associated e-waste, but rather to provide an initial aggregate estimate that highlights the potential scale of the upcoming challenge and Exploring circular economy solutions.
It is by necessity a manual business, projecting the second-order effects of a notoriously fast-moving and unpredictable industry. But someone has to at least make an effort, right? The key is to get it right within the digits, not within the percentages. Are we talking about tens, hundreds of thousands, or even millions of tons of e-waste? Researchers say it's probably near the upper end of that range.
The researchers modeled several low-, medium-, and high-growth scenarios, what kind of computing resources would be needed to support them, and for how long. It has become Their basic finding is that waste will increase by a factor of 1,000 in 2023.
“Our results show that e-waste could rapidly increase from 2.6 thousand tonnes (kt).” [per year] Approximately 4-2.5 million tons (Mt) in 2023 [per year] In 2030,” they write.
Image credit: Wang et al.
Admittedly, using 2023 as a starting indicator is probably a bit misleading. The 2.6 kilotons figure does not include waste because so much computing infrastructure has been deployed in the past two years. This will significantly lower your starting numbers.
But in another sense, this indicator is very real and accurate. After all, these are the approximate volumes of e-waste before and after the generative AI boom. As this first large-scale infrastructure reaches the end of its useful life in the coming years, the amount of waste will increase exponentially.
There are various ways to mitigate this, which the researchers outline (again, only in broad strokes). For example, servers at the end of their useful life can be downcycled rather than retired, and components such as communications and power can also be reused. Software and efficiency improvements may also occur, extending the useful life of a given chip generation or GPU type. Interestingly, they prefer to update to the latest chips as soon as possible. Otherwise, companies would have to buy two slower GPUs to do the job of a high-end GPU, for example, doubling (and possibly accelerating) the resulting waste.
These mitigations have the potential to reduce waste loads by 16-86%. This is obviously quite a range. However, the problem is not so much uncertainty about effectiveness as it is uncertainty about whether and to what extent these measures will be adopted. If every H100 had a second life on a low-cost inference server at some university, the computation would expand significantly. Even if only 1 in 10 people get that treatment, that's not that many.
That is, in their estimation, achieving lower rather than higher levels of waste is a choice, not a necessity. You can read the full study here.