GridFlexDC: Intelligent power system optimisation for AI data centres
Market domain
• Power system operators procure frequency-response services to stabilise the grid
• Requires power flexibility delivered in less than a second, within a tight error tolerance, up to a preagreed MW scale power capacity
Opportunity
• Provide accurate sub-second power adjustment by power capping of aggregated GPU’s in AI Data centres
• Provide frequency response with low activation rates & durations, e.g., <1% of power capacity activated for Great Britain in 2024
Novel Solution
• AI-based fast time-varying GPU power capping to deliver frequency response on top of the efficiency and lifetime improvement
Applications: Data centre power, energy management/control, power system flexibility services
| Features | Benefits |
|---|---|
| Static GPU power capping is widely used to optimise GPU lifetime, efficiency and computing speed | Increases GPU lifetime by 0.7% |
| Oxford AI-based approach makes intelligent power capping time-varying; enables a business-grade grid frequency service | Increases FLOPS/W efficiency by 0.3% |
| Earning revenue with almost zero computing delay | Incurs a negligible computing delay of 0.08% |
| Wider applications include provision of mandatory demand response and improved resilience when data centre microgrid operates independently of the supply grid | 1 GW H100 data centre can provide 500 MW of frequency response and additionally earn £17m/year in Great Britain, €54m/year in Germany, and $19m/year in the US |
| End users include AI companies; data centre operators; data centre energy infrastructure providers; and data centre energy management software providers |
Available For
- Consulting
- Licensing