Examining Energy Consumption in the AI Industry

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Oct 12 2023

In a recent publication in the journal Joule, researchers explored the electricity consumption of artificial intelligence (AI) technologies and its potential impact on global data center energy use.

In 2022 and 2023, artificial intelligence (AI) saw rapid growth, with major players such as Microsoft and Alphabet increasing their support. OpenAI's Chat Generative Pre-Trained Transformers (ChatGPT) achieved remarkable success, garnering 100 million users in two months, prompting Alphabet and Microsoft to introduce their chatbots, Bard and Bing Chat. This expansion raised concerns about AI's energy consumption and environmental impact. Data centers traditionally account for one percent of global electricity consumption, excluding cryptocurrency mining, and there's worry that AI's demands could increase this.

AI and Energy Consumption

AI encompasses a range of methods and technologies that empower machines to exhibit intelligent behavior. Generative AI tools, including ChatGPT and OpenAI's DALL-E, are used for creating new content. Both tools utilize natural language processing and undergo a two-phase process. The initial phase, which is commonly acknowledged as the most energy-intensive, has been a focal point in AI sustainability research.

During this stage, extensive datasets are fed into AI models, and their parameters are fine-tuned to match the predicted output with the desired target. This strategy entails instructing models to predict words or sentences by considering context, especially when dealing with extensive language models such as GPT-3. Models such as GPT-3, Open Pre-trained Transformer (OPT), and Gopher are known to consume a significant amount of electrical power during their training, with the precise amount fluctuating.

Once these models are trained, they are deployed in a production environment and transition into the inference phase. In this phase, they generate outputs based on unseen data. While the environmental impact of the inference phase has received less attention, it should not be underestimated. Some indications suggest that the energy demand during inference may exceed that of the training phase. Google has reported that a significant portion of AI-related energy consumption is attributed to inference.

Nonetheless, contrasting data indicates that certain AI models exhibit significantly reduced energy consumption during inference when compared to the training phase. Various elements impact this proportion, encompassing the frequency of retraining and the delicate balance between model performance and energy expenditure.

AI's 2023 Landscape: Demand and Power Needs

In the year 2023, the AI industry witnessed a phenomenal surge, sparking an ever-increasing appetite for AI chips. By August of that same year, NVIDIA, a prominent player in chip manufacturing, proudly declared an unprecedented revenue of $13.5 billion for the second quarter ending in July 2023. Particularly noteworthy was the data center segment, which experienced a remarkable 141 percent upswing compared to the preceding quarter, a clear indicator of the burgeoning demand for AI-related products. This growing hunger for AI solutions carries with it the potential for a substantial escalation in the energy consumption associated with AI.

Consider a scenario where generative AI, such as ChatGPT, becomes an integral component of every Google search. Projections suggest that such an integration would necessitate a staggering number of NVIDIA servers, resulting in an astronomical surge in electricity usage. The potential ramifications on Google's overall electricity consumption are noteworthy, especially since AI already accounted for 10–15 percent of their energy usage back in 2021. It is essential to recognize that, while this worst-case scenario is a conceivable outcome, it is improbable to materialize rapidly due to resource constraints and related financial considerations.

For a more pragmatic assessment of AI-driven electricity consumption, one can examine NVIDIA's sales and its dominant position in the market. Nevertheless, potential bottlenecks in the supply chain for AI servers remain a consideration. Innovations in model architectures and algorithms do offer the prospect of mitigating the electricity demands of AI. However, it is essential to remain cognizant of the rebound effect, reminiscent of Jevons' paradox, which could come into play.

Furthermore, the increasing trend of repurposing graphical processing units (GPUs) from cryptocurrency mining for AI tasks substantially influences the energy landscape. As model efficiency advances, it introduces a dynamic that influences the trade-off between performance and electricity costs, underscoring the need to strike a balance between efficiency gains and performance enhancements.

Predicting the future trajectory of AI-related electricity consumption remains a formidable challenge. It is imperative to adopt a cautious and balanced approach, steering clear of overly optimistic or pessimistic outlooks. While integrating AI into applications such as Google Search may indeed augment electricity consumption, the constraints of resources and the rebound effect may temper its growth.

Efficiency improvements could lead to heightened demand for AI, necessitating a careful equilibrium between gains and overall resource utilization. Developers should conscientiously weigh the necessity of AI in various applications, and regulators might play a role in enhancing transparency through environmental disclosure requirements, thereby enabling a more profound understanding of the environmental costs associated with this technology.

Conclusion

In summary, researchers explored the energy consumption of AI-driven tools. The study delved into the training and inference phases, highlighting that certain AI models consume significantly less energy during inference compared to the training phase. The relative energy consumption between these phases remains an open question, necessitating further research.