How generative AI learns and creates using GPUs

Ultimately, it is the software models that underpin generative AI that are driving change in the data center. Graphics processing units (GPUs) may be the power-hungry devices that enable effective AI training, but it is innovations in software that have triggered recent hype and investment. In this report, Uptime describes how neural networks — the algorithms behind generative AI — work at a granular level. This will help operators to better understand how to support AI; and where and how to site training and inference systems. This report also provides a brief overview of related topics that will be the subject of future research, such as the implications for computational intensity, IT budgets, data volume and sovereignty.

KEY POINTS

• Many generative AI technologies are based on neural networks. A neural network is a computational model inspired by the human brain that consists of interconnected layers of simple code scripts called neurons.
• Each neuron performs simple mathematics based on tuned parameters called weights, and inputs from data or other neurons. The network’s intelligence comes from its ability to learn the values of these weights, allowing the model to make reasonably accurate predictions and classifications.
• GPUs are ideal for training neural networks because they can simultaneously perform simple mathematics (functions of the neurons) on many data points (which are the weights and inputs).
• GPU power demands come from the many cores being operated in parallel and the rapid movement of data. Larger AI models require more parallelism and more data to train the model in a practical timeframe. AI application-specific integrated circuits (ASICs) may become more popular due to their low cost and small power profile.