Is Your Data Center Facility AI-Ready?
Artificial intelligence (AI) is revolutionizing how we design and build data centers. In simple terms, every data center is now becoming an AI data center.
Traditional data centers are typically designed to support lower power densities, commonly ranging from 6-15kW per rack. These power densities are suitable for standard IT equipment such as servers, storage devices and networking equipment.
However, with the deployment of AI workloads, especially those involving high-performance computing (HPC) or deep learning, the power requirements per rack can be significantly higher, often ranging from 20-50 kW or even more. Training complex neural networks, for instance, requires substantial computational power, resulting in considerable heat generation.
AI's influence on data centers necessitates rethinking data center designs to accommodate these higher power requirements and address the unique challenges posed by AI workloads, including:
Cooling infrastructure
Traditional room cooling systems, designed for lower power densities, typically handle up to 15-20 kW per rack. To ensure effective heat dissipation and temperature control, upgrading or redesigning the cooling infrastructure becomes crucial for maintaining optimal performance and preventing overheating.
Specialized cooling solutions offer various ways to handle high-density racks:
- InRow cooling units achieve higher cooling capacities, reaching up to 50 kW per rack.
- Rear door cooling systems provide even more cooling power, ranging from 50-70 kW per rack.
- Additionally, liquid cooling technologies go beyond air cooling, allowing for scalability. These options include direct liquid-to-chip, chassis or immersion cooling, which can complement existing systems.
Power distribution
Most data centers currently utilize 208V 3-phase power distribution, which can handle power densities of up to 17 kW at 60 amps or 28 kW at 100 amps per rack. However, when dealing with high-density racks, this 208V distribution may not be the most efficient choice for power delivery.
Higher power densities necessitate robust power distribution systems within the data center to ensure reliable and efficient electricity supply to the racks. Achieving this may involve upgrading electrical circuits and distribution panels. As a solution, many data centers are now adopting 415V power distribution. This approach allows for higher power capacities per rack power distribution unit (PDU), providing up to 34 kW at 60 amps or 57 kW at 100 amps per rack PDU.
Space utilization
Higher power densities often mean fewer racks can be accommodated within the same physical space. Data center layouts may need to be reconfigured to optimize space utilization while accommodating the increased power requirements.
Equipment compatibility
Existing racks, PDUs and other infrastructure components may need to be replaced or upgraded to support the higher power densities required by AI deployments.
Cost considerations
Implementing the necessary upgrades to support higher power densities can involve significant costs, including expenses related to equipment purchases, infrastructure modifications and increased energy consumption for cooling.
To address these challenges, some organizations are opting for specialized data center designs tailored specifically for high-density workloads. Others are choosing modular data center solutions that can offer scalability and flexibility to adapt to changing power requirements over time.
Finding the right facilities infrastructure design to support AI workloads and application demands while still allowing for flexibility, scalability and manageability — and balancing cost and efficiency — requires an understanding of the current infrastructure challenges as well as the organization's IT goals and requirements. As AI continues to demand more from data centers, IT leaders must rethink infrastructure, incorporate advanced cooling and power distribution technologies, and upgrade systems to ensure optimal performance, reliability and energy efficiency for AI deployments.