"Smart energy" proposes optimizing energy consumption in AI systems, balancing performance and environmental impact
Advances in computational performance throughout the 20th century were largely driven by the miniaturization of computer components, a trend formalized in 1975 by Intel co-founder Gordon Moore. His prediction β known as Moore's Law β stated that the number of transistors in computer chips would approximately double every two years [1]. However, the pace of semiconductor miniaturization is slowing as the theoretical limits of this approach are being reached β there is not much more room "down there." Avenues remain to improve computing performance, especially at the "top" of the computing stack: software, algorithms, and hardware architectures [2].
The new era of computer architecture focuses on creating software-hardware paradigms to optimize applications in specific domains [3]. Artificial intelligence (AI) is a clear example of this approach, as its computational nature demands specialized solutions, such as the optimization of matrix multiplication operations. This co-design approach aims to maintain the accuracy of robust models while minimizing their energy consumption: it is estimated that training a large AI model can consume up to 552 tonnes of COβ [4], equivalent to the lifetime emissions of ten cars (see Table 1). Faced with this reality, energy efficiency becomes essential to reduce operational costs and environmental impacts. Advancing toward "smart energy" means developing technologies that prioritize both performance and sustainability, ensuring that the benefits of AI are accessible and sustainable for all sectors of society.
Table 1. Comparison of artificial intelligence models in terms of number of parameters, training time, energy consumption, and equivalent gross COβ consumption in tonnes (tCO2e). Estimates from [4].
The explosion of artificial intelligence in the last decade was driven largely by the use of GPUs (Graphics Processing Units) and, more recently, TPUs (Tensor Processing Units). Originally developed for graphics rendering in games, GPUs proved extremely efficient at processing parallel operations, such as the matrix multiplications required to train deep neural networks. However, this efficiency came at the cost of high energy consumption. In response, TPUs emerged, designed specifically for AI processing and offering greater energy efficiency than GPUs β but still consuming large amounts of energy for large-scale models [10].
Figure 1. Futuristic representation of a data center generated by OpenAI's ChatGPT artificial intelligence.
Despite advances in hardware design, the energy consumption of data centers (see Figure 1), where AI systems are trained and deployed, continues to grow rapidly. These centers, which house thousands of GPUs and TPUs, already account for about 1% of global electricity consumption, with forecasts of significant increases due to growing demand for AI [11, 12]. The energy consumed by these devices results in increased needs for cooling infrastructure and robust energy sources, amplifying the environmental and economic impact of developing and using AI models.
Smart energy proposes a new way of thinking about energy consumption in the age of artificial intelligence. It is a concept that goes beyond algorithm optimization, extending to the conscious and efficient use of energy during the training and operation of AI systems. Instead of focusing exclusively on increasing model accuracy, smart energy highlights the importance of balancing performance with energy impact. This means developing solutions that treat energy consumption as a valuable resource, generating a kind of "energy credit" for each algorithm.
Historically, the focus was on maximizing AI model accuracy without adequate attention to the amount of energy consumed in the process. However, as models become more complex and computationally demanding, it is crucial to rethink this paradigm. Smart energy seeks to redefine the success of an algorithm, considering not only its effectiveness but also its energy sustainability.
Energy efficiency in artificial intelligence can be achieved through various technologies and software approaches. Energy-optimized algorithms are designed to reduce energy consumption without compromising accuracy. Techniques such as pruning eliminate unnecessary connections in neural networks, reducing the amount of calculations and, consequently, energy consumption [13]. Quantization reduces complexity by using smaller numerical representations, making calculations faster and more efficient [14]. Knowledge distillation allows training a smaller, more efficient model based on a larger model, maintaining performance while significantly reducing computational and energy costs [15, 16]. Additionally, neural architecture search automates the discovery of neural network architectures that are both accurate and energy-efficient β known in the literature as winning lottery tickets. This search identifies the best configurations in an optimized manner, without manual intervention [17, 18].
The development of green hardware is also fundamental to achieving this efficiency. One of the most promising new architectures is in-memory computing, which integrates processing directly with data storage, reducing the need to move data between processing units and memory [19]. Since data movement is the most energy-intensive operation in traditional digital systems, this approach can drastically reduce energy consumption [20]. Quantum computing offers the possibility of processing information exponentially faster on complex problems, such as the optimization of neural networks. Unlike classical computers, which process data sequentially, quantum computers can perform multiple operations in parallel, significantly reducing processing time and energy consumption [21].
The socioeconomic benefits are significant. Energy efficiency reduces operational costs, making technology more accessible to small and medium-sized enterprises and promoting the democratization of AI access in Brazil. This stimulates innovation and competitiveness in the domestic market. Moreover, lower energy consumption contributes to environmental sustainability, aligning with global emissions reduction goals. Smart energy not only drives technological development but also promotes a positive impact on society and the economy, creating opportunities and improving quality of life.
To position Brazil as a leader in energy efficiency in artificial intelligence, it is essential to establish national guidelines promoting research and development focused on this area. This includes the creation of government programs that encourage universities and research institutes to explore innovative solutions in energy-efficient algorithms and hardware. Incentives for industry are equally important. Proposals such as tax breaks for companies investing in sustainable AI technologies and the provision of specific financing lines can encourage the private sector to adopt smart energy practices. Furthermore, public-private partnerships can accelerate the implementation of these technologies, benefiting the economy and society.
Figure 2. President Lula launching the Brazilian Artificial Intelligence Plan at the 5th National Conference on Science and Technology, on July 30, 2024, in BrasΓlia (Photo: Marcelo Camargo).
The Brazilian Artificial Intelligence Plan (see Figure 2) is a perfect platform to begin this transition. Policies that encourage research in energy-efficient algorithms, along with the development of green hardware, will not only reduce environmental impact but also promote the country's technological sovereignty. It is time to act with determination, incorporating smart energy into national guidelines and ensuring that AI is not only powerful, but also sustainable.
After all, the future of artificial intelligence is not only about intelligent systems, but about an intelligent use of the energy that sustains them. And in that future, Brazil can be a protagonist.
[1] G. E. Moore, "Cramming More Components Onto Integrated Circuits," in Proceedings of the IEEE, vol. 86, no. 1, pp. 82-85, Jan. 1998, doi: 10.1109/JPROC.1998.658762.
[2] Charles E. Leiserson et al., "There's plenty of room at the Top: What will drive computer performance after Moore's law?" Science 368, 2020. DOI:10.1126/science.aam9744.
[3] John L. Hennessy and David A. Patterson. 2019. "A new golden age for computer architecture." Commun. ACM 62, 2 (February 2019), 48β60. https://doi.org/10.1145/3282307.
[4] David Patterson, et al. 2021. "Carbon Emissions and Large Neural Network Training." https://arxiv.org/pdf/2104.10350.
[5] Colin Raffel, et al. 2020. "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer." Journal of Machine Learning Research, 21(140), 1-67.
[6] Daniel Adiwardana, et al. 2020. "Towards a Human-like Open-Domain Chatbot."
[7] Dmitry Lepikhin, et al. 2020. GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding.
[8] William Fedus, et al. 2022. "Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity."
[9] Tom B. Brown, et al. 2020. "Language Models are Few-Shot Learners."
[10] G. S. NikoliΔ, et al., "A Survey of Three Types of Processing Units: CPU, GPU and TPU," 2022 57th International Scientific Conference on Information, Communication and Energy Systems and Technologies (ICEST), Ohrid, North Macedonia, 2022, pp. 1-6, doi: 10.1109/ICEST55168.2022.9828625.
[11] D. Patterson et al., "The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink," in Computer, vol. 55, no. 7, pp. 18-28, July 2022, doi: 10.1109/MC.2022.3148714.
[12] de Vries, A. (2023). "The growing energy footprint of artificial intelligence." Joule, 7(10), 2191β2194. doi:10.1016/j.joule.2023.09.004.
[13] Blalock, D., Gonzalez Ortiz, J. J., Frankle, J., & Guttag, J. (2020). "What is the state of neural network pruning?" Proceedings of machine learning and systems, 2, 129-146.
[14] Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R., & Bengio, Y. (2018). "Quantized neural networks: Training neural networks with low precision weights and activations." Journal of Machine Learning Research, 18(187), 1-30.
[15] Gou, J., Yu, B., Maybank, S. J., & Tao, D. (2021). "Knowledge distillation: A survey." International Journal of Computer Vision, 129(6), 1789-1819.
[16] Geoffrey Hinton, Oriol Vinyals, & Jeffrey Dean (2015). "Distilling the Knowledge in a Neural Network." In NIPS Deep Learning and Representation Learning Workshop.
[17] Jonathan Frankle, & Michael Carbin (2019). "The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks." In the International Conference on Learning Representations.
[18] Ren, P., Xiao, Y., Chang, X., Huang, P. Y., Li, Z., Chen, X., & Wang, X. (2021). "A comprehensive survey of neural architecture search: Challenges and solutions." ACM Computing Surveys (CSUR), 54(4), 1-34.
[19] Sebastian, A., Le Gallo, M., Khaddam-Aljameh, R. et al. Memory devices and applications for in-memory computing. Nat. Nanotechnol. 15, 529β544 (2020). https://doi.org/10.1038/s41565-020-0655-z.
[20] M. Horowitz, "1.1 Computing's energy problem (and what we can do about it)," in 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC). San Francisco, CA, USA: IEEE, 2014, pp. 10β14.
[21] Siddhant Garg, & Goutham Ramakrishnan. (2020). Advances in Quantum Deep Learning: An Overview.
This site collects anonymous analytics (IP, browser, approximate location) for content improvement purposes.
Comments
Loading comments...
Leave a comment. No account needed β just your name.