Empromptu AI has launched a new software solution, Grid Guard, designed to help AI data centres manage power volatility caused by simultaneous graphics processing unit (GPU) workloads.
The system aims to address the challenge of rapid swings in electricity demand, which can reach tens of megawatts in milliseconds when thousands of GPUs perform operations together.
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
Such spikes have the potential to damage power generators and force data centre operators to invest in additional, often costly, infrastructure.
Grid Guard operates by staggering GPU workloads in very short intervals of around 50 to 200 milliseconds, transforming abrupt power spikes into more manageable ramps.
According to initial deployments, the approach has led to an average 80% reduction in power volatility, reportedly without significant impact on the performance of AI workloads.
Empromptu AI has confirmed that the solution is being prepared for a live deployment in collaboration with a major power company.
The software forecasts imminent changes in GPU workload power demand, notifies operators before the power swings reach the electrical system, and phases workloads to smooth out demand.
Grid Guard can reschedule workloads, divide large input data into smaller segments, and reuse computational results to avoid repeated spikes.
The system also sends advance signals to batteries, generators, and power management systems, preparing them for expected changes.
In addition, it quantifies volatility, linking economic signals to data centre load behaviour, and making the cost of inefficient load patterns transparent.
Current industry practices often rely on hardware solutions such as large battery arrays and extra generators to absorb and manage power spikes, leading to increased capital expenditure.
Previous data centre deployments have experienced extensive equipment damage attributed to synchronised GPU usage.
Grid Guard addresses these issues in software, acting before spikes reach the power infrastructure and allowing the solution to adapt over time as it collects operational data.
The company’s platform, initially developed to support enterprise AI model training, supports real-time system optimisation and feedback loops.
