Chatbots like ChatGPT are always drinking, In the course of a typical 20-50 question long conversation they can down upto 500ml of water to cool their servers. They also have a massive carbon footprint-a result of their insatiable energy hunger. Richa Gandhi looks at the ecological impact of Ai’s growing use, and some measures to mitigate it.
500ml
- ChatGPT Needs to “drink” up to 500ml of clean fresh water for a simple conversa- tion of roughly 20-50 questions and answers
Just like organic life, artificial intelligence (AI) cannot exist without water. It uses water directly to cool its massive server rooms, and indirectly at the power station that produce electricity for those servers. The total water consumption of AI is called its ‘water footprint’.
Popular new AI tools like ChatGPT and BARD fall in the category of ‘large language models’ and have a huge water footprint. These models are trained on massive language datasets that are hosed on stacks of energy-hungry servers. Their operations produce of a lot of head, and even if the centers are located in a cold climate, cooling becomes necessary.
Servers work best at 10-27 degrees Celsius, and to maintain this temperature range server farms employ large cooling towers. For every unit (kilo- watt-hour) of electricity consumed by the servers, cooling towers use a gallon (3.8 liters) of water.
How Cooling Towers Work
Cooling towers work on the same principle as traditional room coolers. When water evaporates, it absorbs heat from its surroundings and reduces the ambient temperature. The water vapors rises inside the cooling tower and is released into the atmosphere. As a result, the water used by data centers is lost and cannot be recycled.
This is doubly problematic because cooling towers at data centers can use only clean fresh water. Say, from rivers and lakes. Seawater is not an option because its high salt content would cause corrosion, damaging sensitive equipment at the data center.
Carbon/Water Trade-Off
Data centers located in countries such as Sweden and Finland use less water because of the naturally cooler conditions. But in the Asia-Pacific region, where a lot of the AI action is now concentrated, higher ambient temperatures push up the need for water.
A research paper titled “Uncovering and Addressing the Secret Water Foot- print of AI Models” points out there’s often a trade-off between carbon efficiency and water efficiency. You can generate more solar energy to run servers in the afternoon (so, smaller carbon footprint), but then you need more water for cooling as it is the hottest time of the day.
“AI model developers may want to train their models during the noon time when solar energy is more abundant, but this is also the hottest time of day that leads to the worst water efficiency,”
the research paper’s authors write, citing the example of LaMDA’s training in sun-drenched Nevada. LaMDA is Google’s AI-based conversation technology. In other words, adopting renewable energy can sometimes come in the way of water conservation. The challenge, then, is to find a way to balance carbon and water efficiencies through new approaches to sustainable AI.
What Companies Are Doing
Most AI companies have pledged to make their systems sustainable by 2030. One solution, the paper’s authors say, could be to run AI model training in different locations at different points in time.
Microsoft says its data centers in Phoenix, Arizona, which hosted the training of GPT-3 and its advanced version ChatGPT4, saved water by using outdoor air to chill servers for most of the year. They otherwise cool through direct evaporation, which uses a fraction of the water required by other, more traditional, water-based cooling systems like cooling towers.
Microsoft further plans to save a mil- lion liters of water daily by switching from conventional energy to solar energy from the “Sun Streams 2 Solar Project,” operated by its local partner Long road Energy Google, meanwhile, uses a mix of air cooling, water cooling, refrigerants, or some combination of them, to reduce its water data-driven approach to local hydrology, topography, energy, and emissions issues.
Source : Times of India
