Will We See More Energy Efficient AI Systems In Our Lifetime?
Artificial Intelligence (AI) is significantly impacting energy consumption globally. According to the World Economic Forum, AI-driven data centers are consuming more electricity than countries like South Africa and Indonesia. Energy use is projected to…
Artificial Intelligence (AI) is significantly impacting energy consumption globally. According to the World Economic Forum, AI-driven data centers are consuming more electricity than countries like South Africa and Indonesia. Energy use is projected to rise from 260 terawatt hours in 2024 to 500 terawatt hours by 2027.
A single query in OpenAI's ChatGPT uses 2.9Wh of electricity, which is approximately ten times the consumption of a Google search. With millions of queries daily, the energy demand is substantial.
During the 2024 World Economic Forum in Davos, Sam Altman emphasized the need for an energy breakthrough to accommodate the future growth of AI, which will require more power than anticipated.
Shifting AI processing from large data centers to individual devices offers a potential solution. The World Economic Forum suggests that on-device AI can reduce energy consumption per AI task by 100 to 1,000 times compared to cloud-based systems. This approach minimizes the need for constant data transmission between devices and servers, decreasing overall power usage.
Startups like Groq, DeepSeek, and DeepX are developing chips optimized for energy efficiency rather than computing power. This direction is seen as promising for addressing AI's energy demands.
Governmental actions reflect these concerns as well. For instance, Singapore halted the approval of new data centers between 2019 and 2022 due to energy constraints, highlighting the influence of electricity limitations on technological growth. Additionally, a global energy credit trading system has been proposed, where companies using low-power AI could benefit financially.
According to the World Economic Forum, AI-driven data centers are consuming more electricity than countries like South Africa and Indonesia.
The UK is focusing on decentralized AI, with Nottingham Trent University leading the TinyML UK Network, supported by UK Research and Innovation. The initiative involves collaboration with the University of Southampton and Imperial College London.
TinyML enables machine learning on small, low-power devices rather than large server farms. This allows AI to operate on sensors, wearables, and embedded systems, providing real-time responses and maintaining functionality without internet access.
Professor Eiman Kanjo from Nottingham Trent University highlights the acceleration of AI adoption alongside concerns about energy, cost, infrastructure, privacy, and sustainability. The TinyML UK Network aims to foster low-energy, privacy-preserving AI systems and promote collaboration across various sectors.
TinyML applications include livestock monitors and personal safety devices, which enhance privacy and reduce energy consumption by processing data locally.
The trajectory towards more efficient AI systems is evident as data center energy consumption increases, prompting regulatory limits. Concurrently, advancements in chip design and academic research are focusing on reducing power usage per task.
On-device AI can potentially decrease energy consumption per task by up to 1,000 times. TinyML demonstrates that small models can perform real-world tasks without extensive infrastructure. Support from UK Research and Innovation is propelling this work forward.
It is anticipated that AI will evolve in two parallel directions: large data centers will continue training advanced models, while smaller systems will operate on everyday devices, potentially enhancing services without escalating electricity costs.
Based on reporting by techround.co.uk.
