Seasons of Green DH and Thirsty Computing

Talking about the weather used to be an innocent topic suitable for small talk. These days, this became a political issue, the existence of snow being populistically framed in order to negate climate change, completely misrepresenting its processes.

Weather all over Germany becomes more extreme: the previously two weeks of ‘tropical’ nights of Mainz are stretching longer and longer from June to August, and beginning of December proper centimeters of snow was covering the vineyards near the Rhine for the first time since I moved here. The days, when access to more than enough water in Germany has been regarded as a given, are more or less gone.
In other parts of the world, the scarcity of sweet water has been a known issue for a long time now – and will grow to expose nearly half of the world’s population to extreme water stress by 2030.

Over-usage of water leads to acidification, as (recently brought to attention by a popular satirical news show) in the example of the over-allocated and thus waning Colorado River, whose water allocation was (mis-)regulated in the Colorado River Compact in 1922. In Germany, too, drought is almost the new norm, monitored by e.g. the Drought Monitor on a day to day basis.

And in summer of 2023, the hottest year worldwide up to this date, news hit that the hottest new pieces of technology, the new must-have in industry and academia alike – Large Language Models and their Chat-components – are literally hot. And thirsty, too.
Multiple media outlets reported in spring and summer this year, that LLMs were responsible for 20 to 30% water increases in companies like Google and Microsoft (exemplarily here or here or here or here).

Microsoft’s GPT-3 alone needed approx. 700,000 litres of freshwater for both electriticy generation and server cooling and roughly a half-a-litre bottle of water to answer 30 to 50 prompts, according to estimations by Li et al (2023).

Estimations range the future need of water for AI (and LLM) in the same capacities of the annual water withdrawals of medium-sized countries: in less then 3 years, by 2027, AI could consume as much as half of the UK’s annual water consumption. These dimensions simply become too big to be ignored further.

But the consumption of ChatGPT & Co in perspective is only another (big) drop on a (very) hot stone. Servers are a necessity of the internet age and thus everywhere. And they need to be cooled. Both academic researchers and journalists have warned about the ecological footprints of server centers in general, and machine learning computations specifically (E.g., Strubbell et al 2019).

In order to satisfy the need for coolness of servers, companies become creative. So where to store them pragmatically? Right, in submarines! The feasability of submarine data centres has been tested notably by Microsoft, having launched Project Natick already in spring 2018 at the coast of the Orkney Island, and retrieved the capsule data server center in September 2020 in the North Sea. Performance evaluation showed similar to better efficiancy than land-based data bases. While this solves problems of overheating and minimises electricity use, submerging servers underwater poses a new set of problems, threatening marine flora and fauna, similar to the problems caused by garbage dumps, sunken objects, and underwater construction sites, as for internet cables or mining operations. Other companies too want to launch first commercial subsea data centres, as near the coast of Washington by Subsea Cloud.

In a completely different approach, initiative have strengthened water’s legal status in order to make it able to literally fight back against abuse: in 2020 Lake Eerie in Toledo, Ohio in the US joined an illustrious (but still small) list of environmental entities recognised as juridic persons, such as Vilcabamba River in Ecuador (suing successfully in 2011 for its own rights to ‘exist’ and ‘maintain itself’) , the Ganges and Yamuna Rivers in India (2017), the Te Urewera National Park in New Zealand (2014), and ‘‘Mother Earth’ as a collective subject of public interest in Bolivia, among other examples.

What does all of this have to do with the Digital Humanities? Short answer: alot!

The long answer: DH, too, is part of a bigger problem. DH is contributing to the world’s carbon footprint and water consumption. Digitization efforts, building digital infrastructures, computation times, among others, are contributing to that. So what can we do as Digitial Humanists to mitigate these effects?

Green DH thinks exactly about that.

In March 2023, I attended the “Greening DH” workshop of the DHd working group of the same name in Trier, where we discussed the Digital Humanities Climate Coalition Toolkit. On hyperlinked webpages, this resource offers loads of recommendations and advise on how to act climate responsibly: from ideas of minimal computing, to strategies how to talk with your institute’s superiors to introduce more climate friendliness in the day-to-day business, to recommendations to rather send links to repositories instead of Mail-attachments; the recommendations cover the macro and the micro. Oftentimes this however seems a bit helpless: Analogous to the famous reprimand to ‘close the tap when brushing your teeth,’ it recommends to simply unplug more.
In the workshop itself, we quickly established the differences between individual and structural responsibilities, how we as researchers can become more environmentally responsible in our project management and planning (from the kind of resource providers to choose, infrastructure cycles, sharing licences for proprietary software, to establishing a data managing plan) to organising and attending conferences (regional and meatless catering, ways of transportation, and the under-1000-km-pledge of Scientists4Future, offering more hybrid formats to reduce travel costs and to enhance inclusivity if done right).

But the elephant in the room remained: what exactly is Digital Humanities-specific about Green DH? Is it already DH to have a research data manegement plan? Then rejoice, people, we are all Digital Humanitists already!

This good news aside – and in all seriousness: The reminder, that everything costs something and that resources should not be wasted but mindfully used, is (should be) universalistic, not DH-specific. But that we should strive for more sustainability in our data endeavours, so that the collected data¹ can be accessed and re-used (e.g., for replication studies) in the long term, gets tackled systematically in Germany trough the National Research Data Infrastructure enterprise (NFDI), and its discipline-specific consortia, like NFDI4Memory hosted at IEG. This of course should only be the start. The fact that having a research data management plan has become a requisite for German third-party funding is a result of the German Research Foundation (DFG) making it a requirement for applications. This heralded a change in research culture and spread in the academic landscape: the IEG, too, released in 2022 a RDM guideline mandatory for all its employed researchers. If institutions like the DFG extend their requirements to include green and sustainable considerations of data, it would scale sustainability on structural level and influence future research projects from their very beginning. And let DH finally contribute to mitigate the effects of overconsumption.

 

Literature

Pengfei Li and Jianyi Yang and Mohammad A. Islam and Shaolei Ren (2023). Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. Pre-Print at https://arxiv.org/abs/2304.03271.

Emma Strubell, Ananya Ganesh, and Andrew McCallum (2019). Energy and policy considerations for deep learning in NLP. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 3645–3650, Florence, Italy, July 2019. Association for Computational Linguistics.

 

---

Cite this article as: Cindarella Petz, “Seasons of Green DH and Thirsty Computing,” in Digital Humanities Lab, 15/12/2023, https://dhlab.hypotheses.org/4273.

---

Featured image: Jules Verne (1893) Autour de la lune. 44 dessins par Emile Bayard et Alphonse de Neuville; gravés par Henrie-Théophile Hildibrand. J. Hetzel Paris, here p. 152. https://gallica.bnf.fr/ark:/12148/bpt6k5797067h/f165.item.

1. Notice the semantic oxymoron? Data is always a capta, as Drucker (2011) has reminded us [↩]

---

---