In the fall of 2024, researchers introduced Perseus, an open-source tool designed to reduce the power consumption of training large language models like OpenAI’s GPT by up to 30%. This is achieved by optimizing the software management of GPU usage during AI training.
Recently, Canadian computer scientists discovered another simple method to significantly reduce the energy demand of data centers. This involves optimizing the processing of network packets, as reported by the University of Waterloo. The current method of processing network packets in data centers is inefficient. A minor change in the order of task execution can reduce power consumption by up to 30%.
Martin Karsten, a computer science professor at the University of Waterloo, compares this to optimizing a production line to avoid unnecessary movements. By restructuring the process of network packet processing, Karsten and his team significantly improved CPU cache usage.
To achieve this, only about 30 lines of Linux code need to be modified. Linux is used in data centers by major tech companies. Companies like Amazon, Google, and Meta are cautious about changes. However, the researchers succeeded in getting their modification included in the latest Linux kernel (6.13). Now, tech companies need only to activate this feature to save several gigawatt hours of energy globally.
Reducing the energy consumption of data centers is crucial. Almost every internet service could significantly decrease the energy demand of individual requests, like Google searches or ChatGPT prompts. The energy demand of server farms is expected to increase, driven by AI advancements in the coming years.
According to a May 2024 forecast by Goldman Sachs, the “AI revolution” could increase data center energy consumption by 160% by 2030. This would mean data centers worldwide would need over 1,000 terawatt hours of energy per year, compared to the current 400 terawatt hours.