Green Computing

Deep Neural Networks (DNNs) are amongst the most intensively and widely used predictive models in Artificial Intelligence (AI). Nonetheless, increased computation speed and memory resources, along with significant energy consumption, are required to achieve the full potentials of DNNs. Just to make an example, the energy usage during ChatGPT’s training has been estimated to be equivalent to that of an American household for over 700 years. This course aims at presenting the main reasons behind this extremely high energy consumption and discusses the latest hardware architectures (i.e., TPU, GPU, Custom Designs) designed to allow a sustainable AI. In addition to that, the course will focus on the trustworthiness of AI from the hardware point of view. Indeed, any unexpected behavior can lead to serious consequences up to the damage of physical persons in the case of autonomous vehicles.

• Introduction to deep learning algorithms (CNN, Transformers)
• Von Neumann computing architecture: why it's not suitable for AI
• Specialized computing architecture: dataflow, systolic array and GPUs
• Energy efficiency for AI and computing architectures
• Trusted AI: the role of hardware
• TP: Introduction to a hardware accelerator platform
• TP: AI model modification to fit the hardware accelerator, benchmarking
• BE/TP: Bibliography work/Project