Can Nvidia’s AI Platform Make Data Centres More Sustainable?
Nvidia’s GB300 NVL72 platform tackles a major challenge in AI training: energy spikes that strain electricity grids. By combining onboard energy storage with smart power management, it eases the burden on data centre infrastructure and helps prevent costly over-provisioning.
Managing Sudden Power Surges from Synchronized GPUs
AI training often involves thousands of GPUs operating in sync, causing abrupt spikes in power demand. Unlike traditional data centre workloads that spread out energy use, simultaneous GPU ramp-up and ramp-down create sharp, sudden shifts. These fluctuations can stress transformers, cause voltage instability, and disrupt other grid users. Traditional power generation systems may take up to 90 minutes to respond to these demand swings, making the problem worse.
Nvidia’s solution introduces a coordinated energy management system across three GPU workload phases: ramp-up, steady state, and ramp-down. Instead of an immediate surge, a power cap function gradually increases GPU power draw during ramp-up. At job completion, a “GPU burn” feature holds power consumption briefly to smooth the drop-off. This controlled approach aligns GPU demand more closely with the grid’s capacity.
Cloud provider CoreWeave recently became the first to adopt the GB300 NVL72 platform, highlighting the growing demand for infrastructure that supports large-scale AI training more efficiently.
Energy Storage at the Rack Level Flattens Power Demand
The GB300 NVL72 packs electrolytic capacitors into each shelf to store energy locally. These capacitors charge during low-demand periods and release power during spikes, smoothing out the power curve seen by the grid. Internal testing by Nvidia shows the new racks reduce peak power spikes by 30% compared to previous models, without sacrificing GPU performance.
Partnering with LITEON Technology, Nvidia redesigned the power shelf to dedicate half its volume to energy storage. Each GPU benefits from 65 joules of storage, managed in real time by an embedded controller that balances charge and discharge cycles. This local energy smoothing reduces stress on the overall power system and improves alternating current (AC) stability.
The system allows granular control through Nvidia’s System Management Interface (SMI) or the Redfish protocol. Operators can adjust settings like idle time before ramp-down and how quickly power levels change, tailoring the energy profile to specific workloads.
Reducing Infrastructure Costs and Improving Efficiency
Data centres traditionally overbuild power infrastructure to handle peak loads, resulting in wasted capacity most of the time. Nvidia’s energy smoothing lets operators size infrastructure closer to average demand instead of maximum peaks. This means more racks can fit into the same power allocation, or total power needs can be lowered.
Importantly, all energy smoothing happens within the data centre rack itself; no energy is fed back into the utility grid. This local approach avoids dependency on external grid coordination and simplifies deployment.
As AI models grow larger and training workloads become more demanding, systems like the GB300 NVL72 offer a practical way to improve sustainability and cost-effectiveness in data centre operations.
For operations professionals looking to deepen their knowledge of AI infrastructure and efficiency, exploring specialized courses can provide valuable insights. Check out Complete AI Training for relevant resources tailored to AI and data centre environments.
Your membership also unlocks:
