Generative AI (Gen-AI) represents a major growth potential for the digital industry, a new stage in digital transformation through its many applications. Unfortunately, by accelerating the growth of digital technology, Gen-AI is contributing to the significant and multiple environmental damage caused by its sector. The question of the sustainability of IT must include this new technology and its applications, by measuring its environmental impact. To best respond to this challenge, we propose various ways of improving the measurement of Gen-AI's environmental impact. Whether using life-cycle analysis methods or direct measurement experiments, we illustrate our methods by studying Stable Diffusion a Gen-AI image generation available as a service. By calculating the full environmental costs of this Gen-AI service from end to end, we broaden our view of the impact of these technologies. We show that Gen-AI, as a service, generates an impact through the use of numerous user terminals and networks. We also show that decarbonizing the sources of electricity for these services will not be enough to solve the problem of their sustainability, due to their consumption of energy and rare metals. This consumption will inevitably raise the question of feasibility in a world of finite resources. We therefore propose our methodology as a means of measuring the impact of Gen-AI in advance. Such estimates will provide valuable data for discussing the sustainability or otherwise of Gen-AI solutions in a more transparent and comprehensive way. We intend to help this discussion by differentiating in our approach between the embodied and operational impacts of Gen-AI. In this way, we can consider the sustainability of models, as we already do for equipment.

Deep Learning (DL) frameworks such as PyTorch and TensorFlow include runtime infrastructures responsible for executing trained models on target hardware, managing memory, data transfers, and multi-accelerator execution, if applicable. Additionally, it is a common practice to deploy pre-trained models on environments distinct from their native development settings. This led to the introduction of interchange formats such as ONNX, which includes its runtime infrastructure, and ONNX Runtime, which work as standard formats that can be used across diverse DL frameworks and languages. Even though these runtime infrastructures have a great impact on inference performance, no previous paper has investigated their energy efficiency. In this study, we monitor the energy consumption and inference time in the runtime infrastructures of three well-known DL frameworks as well as ONNX, using three various DL models. To have nuance in our investigation, we also examine the impact of using different execution providers. We find out that the performance and energy efficiency of DL are difficult to predict. One framework, MXNet, outperforms both PyTorch and TensorFlow for the computer vision models using batch size 1, due to efficient GPU usage and thus low CPU usage. However, batch size 64 makes PyTorch and MXNet practically indistinguishable, while TensorFlow is outperformed consistently. For BERT, PyTorch exhibits the best performance. Converting the models to ONNX usually yields significant performance improvements but the ONNX converted ResNet model with batch size 64 consumes approximately 10% more energy and time than the original PyTorch model.

Discover amazing ML apps made by the community