The immense and rapidly advancing computational demands of AI models could lead to the industry disposing of the e-waste equivalent of greater than 10 billion iPhones per 12 months by 2030, researchers predict.
In an article published within the journal NatureResearchers from the University of Cambridge and the Chinese Academy of Sciences are attempting to predict how much e-waste this growing industry could produce. Their goal shouldn’t be to limit the adoption of the technology, which they are saying is initially promising and possibly inevitable, but to higher prepare the world for the tangible results of its rapid spread.
Energy costs, they explain, have been closely examined because they already play a job.
However, less attention has been paid to the physical materials involved of their life cycle and the waste stream of obsolete electronic devices.
The aim of our study shouldn’t be to accurately predict the quantity of AI servers and associated e-waste, but fairly to offer initial rough estimates that illustrate the potential scale of the challenge ahead and to explore possible circular economy solutions.
It is inevitably an unpredictable business, projecting the secondary consequences of a notoriously fast-moving and unpredictable industry. But someone has to at the very least try, right? It's not about getting it right inside a percentage, but inside an order of magnitude. Are we talking about tens of hundreds of tons of e-waste, a whole bunch of hundreds or thousands and thousands? According to the researchers, it might be on the upper end of this range.
The researchers modeled some low, medium and high growth scenarios, in addition to the kind of computing resources that might be needed to support that growth and the way long it will last. Their basic finding is that waste would increase a thousand-fold by 2023:
“Our results suggest a possible for rapid growth in e-waste from 2.6 thousand tonnes (kt) (per 12 months) in 2023 to around 0.4-2.5 million tonnes (Mt) (per 12 months) in 2030 there,” they write.
Admittedly, using 2023 as a baseline is maybe a bit misleading: since much of the computing infrastructure was deployed within the last two years, the two.6 kilotons figure doesn’t include it as waste. This lowers the initial value significantly.
But in one other sense, the important thing figure is sort of real and accurate: in spite of everything, it’s the approximate amounts of electronic waste before and after the generative AI boom. We will see a pointy increase in waste levels as this primary major infrastructure reaches the top of its life in the following few years.
There are several ways to mitigate this, which the researchers outline (again, only in broad strokes). For example, servers may very well be shut down at the top of their life fairly than thrown away, and components comparable to communications and power is also repurposed. Software and efficiency is also improved, extending the effective lifespan of a selected generation of chips or kind of GPU. Interestingly, they like updating to the most recent chips as quickly as possible, otherwise an organization could have to purchase two slower GPUs to do the work of 1 high-end GPU – doubling (and possibly speeding up) the resulting waste.
These remedies could reduce waste by 16% to 86% – obviously a wide selection. But it shouldn’t be a lot the uncertainty about effectiveness because the uncertainty about whether and to what extent these measures might be taken. If every H100 gets a second life on a low-cost inference server somewhere at a university, that shifts the bill significantly; If just one in ten receives this treatment, this shouldn’t be a lot the case.
This implies that of their estimation, reaching the low end of waste versus the high end is a alternative, not an inevitability. You can read the complete study here.
