Environmental Costs and Opportunities of AI (1/2)
 |
Does Artificial Intelligence (AI) not produce any waste ?
False
AI does operate on physical infrastructure, such as data centres which produce electronic waste, which often contains hazardous substances, like mercury and lead. AI data centres often use more water to cool electrical components, and some of that water will become wastewater too.
Which resources are needed for AI to operate ?
- Energy
- Water
- Silicon
- Copper
- All of the above
AI hardware relies on many elements like aluminium, silicon and copper (and many more) to build semiconductors and hardware for data centers. Water is used for cooling and energy of course to power training AI models and operating services using them.
AI and machine learning are remarkably effective at spotting data patterns, streamlining operations, and revealing hidden relationships. They address intricate challenges with a speed and efficiency that was previously out of reach. However, this efficiency comes with a cost. While AI is powerful, it’s also resource-intensive and as AI becomes as widespread as computers, its resource demands will continue to grow. The more we rely on AI to streamline business operations, administrative tasks, and daily activities, the more energy and computational resources we’ll need to support it.
Let us see how AI systems have evolved over the past 75 years:
 |
| AI evolution over time Source: Our World in Data |
The chart depicts a dramatic progress in the last decade, reflected on the amount of computation (and availability of training data) used to train the largest AI systems. With more people using digital services, we are creating a huge amount of data, from text, images, videos, voice, and other media for AI to be trained on and learn from. As more data is added, the machine learning systems get progressively better at dealing with the data, and drawing information and conclusions from it.
Machine Learning (ML): A branch of artificial intelligence (AI) focused on developing computer systems that can learn and adapt without following explicit instructions. ML uses algorithms and statistical models to analyze and draw inferences from patterns in data. Over time, ML systems improve their performance and accuracy through experience and exposure to more data.
Some examples include Netflix or Spotify as they analyze user preferences and viewing/listening history to suggest content that aligns with individual interests. The algorithms improve as more data is collected over time.
 |
| ChatGPT interface Source: OpenAI |
The spread of generative AI in homes is accelerating at a pace far beyond that of earlier digital technologies such as the internet and personal computers. Because training AI models and operating related applications demand significant computing power, they consume more energy than their predecessors. That’s why organizations like the International Energy Agency (IEA) predict that the AI boom will be followed by a sharp rise in energy demand, particularly for local renewable energy sources to power data centers.
 |
| Household adoption of Generative AI (GAI) Source: INCO Group |
Beyond Chatbots
 |
| A video generated from a text prompt using OpenAI's Sora AI model Source: Wikipedia |
The question now arises:
What resources are required to create such video ?
How much energy does an AI model needs to generate it ?
Physical Presence of AI
Let's begin by understanding what's behind an AI tool like the ChatGPT chatbot.
Conversations around AI’s energy use tend to emphasize electricity, largely because it’s straightforward to quantify and compare. Yet this narrow focus misses other critical dimensions like water usage, land and the ecological impact of waste produced during infrastructure development, power generation, and cooling processes. Keep in mind that extraction of minerals for renewable energy technologies and ICT equipment plays a crucial role.
Data Centers
Buildings used to house computer systems and associated components, such as telecommunications and data storage systems. To build a new data center, land has to be used and adapted, often close to a water source. As with all industrial buildings, it will require materials, electricity and will create waste both during construction and operation.
 |
| Aerial view of the Google Data Center Source: http://chaddavis.photography/ |
In Ireland, where taxes for data center operators are the lowest, data centers could account for nearly one-third of the country’s electricity demand by 2026 (Madaline Dunn, 2023). In 2022, data centers in Ireland used 5.3 TWh (terawatt-hours) of electricity, making up 17% of the country’s total consumption.
One of the main concerns with Artificial Intelligence and the environment is that data centers need a constant, stable supply of energy to function. This is essential not just for running the servers, but also for powering equipment that ensures efficiency and prevents downtime. That's why the energy capacity of a data center is highly important. The size, design, and energy efficiency of data centers can vary, as can the sources of their energy.
Large 'hyperscale' data centers, for example, can house more than 5,000 servers, span over 1,000 square meters, and require at least 40 megawatts (MW) of energy capacity. The largest data center in the world, the China Telecom Data Center in Hohhot, China, covers 994,062 square meters and has a 150 MW energy capacity. To put that into perspective, a solar farm with the same energy output could power between 30,000 and 150,000 homes.
Computers and Servers
Servers, with their core components such as Central Processing Units (CPUs), Graphic Processing Units (GPUs), memory (RAM), hard drives, and fans, all need electrical power to operate. AI data centers operate with higher load intensity and often require more energy and cooling in comparison to data centers that host websites and cloud services.
 |
| CSIRO Supercomputer cluster with traditional CPUs but with more powerful GPUs. Source: CSIRO |
GPUs and CPUs
Every AI system requires specialized processing units to train the model and perform tasks later (referred to as inference). Large language models (LLMs) are trained and run on thousands of GPUs, similar to the graphics cards needed for modern video games.
Producing CPUs and GPUs involves mining raw materials such as silicon, copper, and aluminum. Additionally, precious metals like gold and silver are used to create the wires for these components.
 |
| View of the 4 high-performance graphics cards NVIDIA H100 that are used in training AI models Source: 极客湾Geekerwan |
GPUs (Graphics Processing Units): Specialized processors designed to handle parallel computations. GPUs excel in training deep learning models due to their ability to process large datasets simultaneously. GPU-based servers provide higher performance for machine learning, neural networks, and deep learning tasks.
Server: A server is a computer designed to provide services, data, or resources to other computers (called clients) over a network. It’s like a digital host that listens for requests and responds with the right information or action. They’re built for reliability, speed, and continuous operation, often running 24/7 in data centers. AI servers, on the other hand, are like the high-performance athletes of the server world (built for speed, power, and complexity).
 |
| Water cooling system Source: U.S Department of Energy (DOE) |
Power and water use in data centers are closely linked; cooling systems that require less power typically consume more water, and vice versa. Operators must carefully choose cooling technologies that balance these resources based on their site’s needs, the availability of resources, cost, and local community demands.
A key metric for measuring water consumption in data centers is Water Usage Effectiveness (WUE), calculated as:
WUE=Actual Site Water Usage / Information Technology Equipment Energy
WUE = Liters per kW-h (L/kWh)
Networking Equipment
Any IT infrastructure relies on a variety of networking devices, such as routers, switches, and firewalls, to maintain connectivity both within the data center and with the outside world. These devices, like CPUs and GPUs, also require power to function and are made from various raw materials and rare metals.
 |
| Source: Pexels |
Key Takeaways
- Servers, with their core components such as the CPUs, GPUs, memory (RAM), hard drives, and fans, all need electrical power to operate. They can be responsible for 50% to 70% of the total power consumption.
- While water reuse and recycling is common practice for many data centres, other processes like mining, construction and chemical cooling will cause water pollution.
- The extraction of minerals for both renewable energy technologies and ICT equipment, plays a crucial role. Critical materials such as silicon, cobalt, and rare earth elements are geographically dispersed and rarely co-located, which complicates their sourcing, presenting significant geopolitical and environmental challenges.
Comments
Post a Comment