AI load and chiller systems: key considerations

A defining characteristic of today’s generative AI training workloads is their choppy load profile. While many other workloads exhibit highly varied activity levels, training compute is different because it can produce dramatic power fluctuations at scale — ranging from hundreds of kilowatts to several megawatts.

Some of these fluctuations are almost instantaneous and frequent — every few seconds, a quick surge or fall of currents in the electrical system takes place within milliseconds. These result from silicon-switching and synchronization events that make many GPUs work in unison. Even between these “macro” events, significant oscillations can occur at millisecond frequency. These high-frequency events primarily stress electrical components.