Responsible AI: Balancing CO2 Emissions and Energy Efficiency for a Sustainable Future

Meisong Yan and Silviu Livescu

Pandemic accelerated digitalization not only in oilfield but also crossing all other industries. While Artificial Intelligence (AI) becomes a new norm to pursue productivity increases and even is viewed as the next industrial revolution, responsible AI has become the next frontier for new ethical codes and standards to protect against potential risks.?

We all want AI to fulfill its promise to improve our lives. For example, in the healthcare industry, patient’s data could be used for machine learning algorithms to enhance the diagnosis accuracy. However, how should we protect patient’s privacy and security in this process? Where the data should be stored? Who has the access rights? And how the data are used only for using machine learning algorithms? It may be a long journey to solve those challenges in a responsible and ethical way.

In the Energy domain, AI is also considered to improve the energy efficiency using such technologies as smart meters or smart grid. One of applications is Virtual Power Plant (VPP) - providing demand response solutions for the electricity grid. A virtual power plant is a network of decentralized energy resources, such as rooftop solar panels, battery storage systems, and smart thermostats, that can be aggregated and controlled as a single entity to provide grid services. This program allows customers with compatible appliances, such as water heaters and air conditioning units, to participate in demand response programs and provide grid services by using their appliances as part of a virtual power plant. This technology has been deployed by Tesla in South Australia, Next Kraftwerke in Germany, and Tiko in Switzerland.?

However, Enterprise-Grade AI platforms are energy intensive to be operated every day. According to the to the International Energy Agency (IEA), the estimated worldwide electricity consumption of data centers in 2022 ranged from 240 to 340 TWh, accounting for approximately 1-1.3% of the total global electricity demand. The combined electricity usage of major tech companies like Amazon, Microsoft, Google, and Meta surged significantly, more than doubling between 2017 and 2021, reaching around 72 TWh by 2021. The rise of AI presents a substantial growth requirement for data centers, as well as the increased energy consumption.

Also reported by IEA, the carbon dioxide emissions from data centers and data transmission networks utilized for digitalization reached 330 metric tons of CO2 equivalent in 2020. This amount corresponds to approximately 0.6% of the total greenhouse gas (GHG) emissions or 0.9% of the GHG emissions associated with energy consumption. Let’s use the popular Large Language Model (LLM) – ChatGPT as an example. Stanford’s study shows that GPT-3 could generate almost seven times more CO2 footprint than a car’s average emission in lifetime. On one hand, AI systems could optimize the energy utilization and push for better energy efficiency. On the other hand, its operation increases the energy consumption. There are numerous articles about the AI system’s extremely high energy consumption. How could we select a responsible AI system that would balance between its own CO2 emissions and its functionality on energy efficiency enhancement?

A world-wide CO2 emission tracker system for the baseline comparison and future energy efficiency optimization has already been developed. After all, not all electricity is generated equally. In different countries or different regions within the same country borderline, one kwh of electricity contains the same power but has a different CO2 footprint.

To be a responsible citizen, each individual could start reducing the CO2 emissions from our own computer usage: just simply turn off our laptops when we are away. A turned off computer uses at least 65% less energy than a computer left on or idle on a screen saver mode. CodeCarbon has released a Python package on Github that estimates the hardware electricity power consumption (GPU + CPU + RAM) and applies it to the carbon intensity of the region where the computing is conducted. In this way, the coders could track their software carbon footage and develop more environmental-friendly applications.

At enterprise level, using federated learning for large-scale AI solutions for model training is highly recommended for a sustainable technology deployment by Dr. Amita Kapoor and Ms. Sharmistha Chatterjee , the authors of “Platform and Model Design for Responsible AI”.?

Federated learning is a distributed learning framework that enables multi-institutional collaborations on decentralized data with improved protection for each collaborator's data privacy. It’s a middle path between a centralized learning and a decentralized learning.?

The CO2 emission calculator for a Federated Learning would consider the hardware devices, the location of the electricity used, the distribution of the dataset, the number of the rounds to build the global model by the central server, the number of the local epochs to train the model, the number of the active devices in each round, and the internet speed of the used devices.?

In the end, we applaud and enjoy the efficiency and the productivity every technology advancement has brought to us. However, at the early stage of each technology development, unforeseen side-effect could co-exist even it’s against our initial purpose. Responsible AI needs to be problem-solving, transparent, fair, and sustainable while avoiding negative consequences. Let’s develop and deploy technology responsibly to ensure its sustainability in the energy domain.

References:

Responsible AI: https://www.accenture.com/us-en/services/applied-intelligence/ai-ethics-governance

Virtual Power Plants (VPP): https://www.energy.gov/lpo/virtual-power-plants

Turns Electric Water Heaters into Grid Batteries: https://www.canarymedia.com/articles/grid-edge/this-startup-turns-electric-water-heaters-into-grid-batteries

IEA: https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks#overview, retrieved July 29th, 2023

AI’s Growing Carbon Footprint: https://news.climate.columbia.edu/2023/06/09/ais-growing-carbon-footprint/

Keeping Watch on the Carbon Footprint of Machine Learning: https://hellofuture.orange.com/en/keeping-watch-on-the-carbon-footprint-of-machine-learning/

“The AI Index 2023 Annual Report,” AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA, https://aiindex.stanford.edu/report/

CarbonTracker CT2022: https://gml.noaa.gov/ccgg/carbontracker/

CO2 Tracker from NASA: https://climate.nasa.gov/vital-signs/carbon-dioxide/

Estimate and Track Carbon Emissions from Your Computer, Quantify and Analyze Their Impact: https://github.com/mlco2/codecarbon?

Carbon Intensity of Electricity: https://ourworldindata.org/grapher/carbon-intensity-electricity?time=2022

Federated Learning: https://en.wikipedia.org/wiki/Federated_learning

Platform and Model Design for Responsible AI: Design and Build Resilient, Private, Fair, and Transparent Machine Learning Models: https://www.amazon.com/Platform-Model-Design-Responsible-transparent/dp/1803237074/ref=cm_cr_srp_d_product_top?ie=UTF8