Data centers running AI workloads are putting a strain on the electrical grid due to fluctuating power demands. This poses challenges for grid operators, causing spikes or sags in voltage for other grid customers. NVIDIA has introduced a new power supply unit with energy storage that can reduce peak grid demand by up to 30%.

In AI training, thousands of GPUs synchronized to perform the same computation cause power fluctuations at the grid level. Unlike traditional workloads, AI workloads result in abrupt transitions between idle and high-power states. NVIDIA’s power smoothing solution in the GB300 platform addresses these challenges by implementing various mechanisms.

The new power cap feature in the GB300 platform gradually increases GPU power draw at the start of a workload, aligning with grid ramp rates. Energy storage elements in the power shelves charge during low demand and discharge during high demand. A power burn hardware and software algorithm ensures a smooth transition when a workload ends.

Empirical results show significant improvements with the new power supply unit in reducing peak power demand by 30% when training AI models. Energy storage in the power supply unit plays a key role in dampening rapid fluctuations and optimizing power consumption. This innovation allows for more racks within the same power budget in data centers.

Integrating energy storage and load smoothing technologies in data centers can optimize peak power consumption, increase compute density, and reduce operating costs. NVIDIA’s advanced power smoothing solution in the GB300 platform offers a promising solution for data center operators looking to enhance power efficiency and grid stability.

Read more at NVIDIA: How New GB300 NVL72 Features Provide Steady Power for AI