Key pannellist included: 

• Chris Lacey, Principal Architect, Industry HPC & AI, Hewlett Packard Enterprise

 The speaker highlighted the increasing challenges associated with scaling computational resources, particularly in relation to AI's pre-training, post-training, and long-thinking phases, which require massive amounts of data and compute power. This has led to a significant rise in power requirements, particularly for CPUs and GPUs, with projections suggesting industrial-type GPUs will require over 4 kilowatts by 2035. Liquid cooling is identified as the inevitable solution for managing the heat produced by these high-performance components, as air cooling becomes impractical.

The discussion also covered the architectural evolution necessary for future data centres to handle these demands. Modular and disaggregated designs were proposed, where components like CPU, GPU, memory, and storage are housed separately to optimise space, power, and cooling efficiency. This approach mirrors older mainframe architectures but is updated with contemporary technology to meet current needs. Direct liquid cooling (DLC) is emphasised as a more efficient method, improving operational efficiency and reducing energy costs. The speaker illustrated this shift with examples from supercomputing, noting that leading-edge AI platforms are adopting similar energy profiles and cooling strategies.

Finally, the session addressed the sustainability and strategic implications of these changes. Energy Edge data centres located near renewable energy sources or facilities with high energy reuse potential were proposed as a viable solution to the increasing power consumption. Such centres could leverage excess energy from wind farms or solar arrays, making them more sustainable and cost-effective. Additionally, legacy data centres will still play a crucial role, providing low latency, high-performance interconnects and serving as hubs for updating models and data. Efficient management of these resources using AI-driven workflow management and data concierges was recommended to ensure optimal performance and sustainability.

Takeaways

Escalating power demands will reshape data centre design

Advancements in AI require massive computational resources, leading to significant increases in power requirements for CPUs and GPUs. Future data centres will need to adopt modular and disaggregated designs to efficiently manage space, power, and cooling.

Direct liquid cooling is essential for future high-performance computing

Air cooling is becoming impractical for the heat generated by advanced CPUs and GPUs. Direct liquid cooling offers a more efficient method of heat transfer, reducing operational costs and improving performance.

Sustainability will drive the location and design of future data centres

Energy-intensive data centres will be located near renewable energy sources or facilities with high energy reuse potential. This approach will leverage excess energy, making data centres more sustainable and cost-effective.