How is physical artificial intelligence used to optimize data center efficiency?

Physical AI (PAI) in data center power systems uses machine learning for predictive maintenance, energy optimization, load balancing, and physical security. In essence, PAI is being used for data center optimization to support the demands of digital AI (DAI) applications like training large language models (LLMs), running inference for real-time applications, and supporting infrastructure like power-hungry GPUs and memory.

PAI supports data center sustainability using thousands of sensors for temperature, airflow, humidity, and power usage. Plus, predictive analytics and intelligent automation are implemented with data center infrastructure management (DCIM) software to cut costs, boost reliability, and improve sustainability by integrating with smart grids and managing complex AI-driven hardware demands.

PAI plus DCIM enables proactive management and control of data center infrastructure, including cooling systems, power distribution, and robotics for maintenance. Important aspects of PAI plus DCIM include (Figure 1):

Sensor fusion for intelligent cooling. PAI analyzes real-time data on temperature, airflow, humidity, and power consumption to optimize cooling on a micro scale, enabling up to 40% reductions in data center cooling energy requirements.

Workload orchestration and dynamic energy management based on a myriad of factors, including availability of green energy resources, depth of discharge in the data center battery energy storage system (BESS), capabilities of various servers and memory systems, and so on. The result is an optimized power usage effectiveness (PUE) rating that drives increased sustainability, reduced costs, and improved reliability.

Figure 1. Three ways PAI can contribute to improved data center operation. (Image: Nlyte)

Sustainability

In addition to optimizing data center cooling system operation and improving PUE ratings, PAI plus DCIM can optimize systems for capturing and repurposing waste heat for use in secondary applications, particularly in colder climates. AI can predict changes in server workload, allowing heat recovery systems to adjust accordingly and ensure a steady supply of heat to greenhouses or district heating systems.

PAI-supported heat repurposing systems can reduce datacenter operating costs by creating a valuable resource from undesirable heating. In large data centers that can enjoy the most benefit from heat repurposing, heat generation can vary significantly, and PAI-based systems can be especially useful.

PAI can help data centers optimize the integration of green energy sources like solar power by predicting energy needs and controlling power switching, energy distribution reconfiguration, and energy storage operations to manage integration into the local grid.

Scalability and security

PAI plus DCIM thermal management can also accommodate the scalability and security needs of cloud data centers. Scalability is supported by adaptive thermal management that morphs as needed to support varying computing demands and expanding computing infrastructure.

Security is a little more complex. Google’s approach is instructive and includes at least eight key elements, some of which include (Figure 2):

The AI estimates the confidence that recommended actions will produce a positive outcome. Recommendations with low confidence of providing a benefit are eliminated and not considered further.
Automated two-layer verification is implemented for actions that pass confidence testing. The recommended actions are first compared with a list of general safety requirements for all data centers operated by Google. Instructions that meet those requirements are delivered to individual data centers, where it’s verified that they also satisfy the requirements of that specific data center operation.
Ultimately, data center operators make the final determination and can terminate AI control mode at any time, and the system will seamlessly transfer to more conventional automation controls without using AI.

Figure 2. Eight examples of PAI and DCIM data center safety tools used by Google. (Image: Google)

Summary

PAI extends existing AI-based data center energy management tools by increasing the level of automation and making them more proactive. PAI relies on massive deployments of sensors to support automated energy management and improved PUEs. It can also support more sustainable operations by integrating the data center into district heating operations and integrating green energy resources into the data center.