Erratic power profiles of AI clusters: the root causes

Ever since ChatGPT-3 arrived with a bang in late 2022, generative AI has dominated discussions among those working in the data center industry. Most of the attention has been directed at the high power density of hardware for AI, as well as the sheer size of the infrastructure needed to run some of these workloads. However, there is a much less publicized issue around generative AI training: its potentially erratic power profile.

In discussions with vendors and data center operators, Uptime Intelligence has come across several independent reports of AI training clusters producing big amplitude power swings. What makes these power events unusual is their scale — they are large enough to create rapid and large variations in the electrical load on UPS systems, the grid or engine generators.