- within Immigration topic(s)
In our recent commentary on generative artificial intelligence (AI), available here, we explored the growing energy demands of data centres. Generative AI refers to algorithms capable of producing various types of content, including text, images, videos, and other data, in response to user queries. Well-known examples include Gemini, ChatGPT, AlphaCode, Midjourney, and DALL-E. These models rely on data centres that house the processing infrastructure required to train and operate them. Analysts estimate that the power demand of these centres will increase by 160% by 2030.
Given this rising energy demand, tech companies are exploring alternative energy sources. Our thoughts on the potential use of nuclear power to meet the energy needs of data centres can be found here. Additionally, significant recent news coverage seen here, has focused on the water usage in the cooling systems used in data centres. Beyond sourcing low-carbon energy and closed-loop cooling systems to conserve water, what other optimisations can be made to reduce the environmental impact of generative AI?
Monica Batchelder, Chief Sustainability Officer at HPE, addresses this question in her recent article, focusing on strategies to improve the energy efficiency of AI models themselves thereby reducing data centre usage and in turn energy demands and/or cooling water usage. Patents have a role to play in protecting these improvements.
Data Cleaning
Typically, generative AI involves two stages: training—where models learn from vast datasets, and inference—where trained models are applied to solve real-world problems. The quality of training data does not only influence the model's performance, but also affects energy efficiency. As Batchelder notes, "Models trained on irrelevant, unstructured, or low-quality data not only underperform, but also waste valuable compute resources." Improving data quality through cleaning techniques, such as removing duplicates, deleting empty records, normalising data, and applying validation checks, can help ensure consistency, completeness, and accuracy, ultimately enhancing model efficiency.
An exemplary patent application related to this innovation is EP4128273A1. This patent application describes using a first AI model trained on cleaning and correcting data. The first AI model so trained then cleans and corrects data before it is used to train a second AI model. While the use of two AI models may appear to consume more computational resources, a model trained on clean, validated data requires fewer computational cycles to reach an acceptable level of accuracy. Depending on data set sizes, the use of two models can still result in an overall reduction in processing resources and accordingly energy use.
Quantisation and Guardrails
Further energy savings can be achieved through techniques like quantisation, which reduces the precision of data used in large language models (LLMs). By converting high-precision weights and activations to lower precision, computational loads and energy consumption can be significantly reduced, especially in applications where high precision is not essential. Additionally, guardrails can be implemented to redirect simple queries to smaller, less resource-intensive models.
Another patent application, EP21798231, relates precisely to this type of innovation. This invention relates to dynamic quantization based on the content of the layer input. This allows the AI system to compress information to maximise energy efficiency without significantly affecting the model's performance.
Hardware Optimisation
Hardware optimisation also plays a critical role. Each hardware component should be selected for maximum efficiency – delivering high performance while consuming less energy and generating minimal heat. This could include advanced cooling systems such as liquid cooling or cooling equalisers, as well as specialised processors designed for AI workloads and large-scale data processing.
EP3895076A1 is an example of innovation in this sub-sector. This invention utilises power consumption data as inputs to an AI algorithm. The result is optimised parameters which reduce power consumption.
Summary
Ultimately, the type of energy used, the design of efficient and specialised hardware, and improvements in software and data quality must all work together to address the energy demands of generative AI. As Monica Batchelder puts it:
"By broadening our field of vision to evaluate the overall AI ecosystem, rather than the underlying hardware alone, we stand a better chance of addressing AI's sustainability challenges..."
An exemplified by the various patent applications identified, protecting these innovations will lead to a reduction in energy and water usage. It will also allow innovators in this space to leverage their knowledge into a commercial and competitive advantage.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.