Executive Summary
- Duncan Singer, Technical Manager – Advanced Products at Excool, talks to DCi about the changing relationship between power resilience and cooling infrastructure.
- This interview explores the impact of AI, increasing rack densities and changing cooling requirements on data centre resilience.
- And, why designing for resilience at block level, alongside scalable and decoupled cooling systems, can help reduce costs and improve energy efficiency.
As AI workloads and rack densities continue to increase, data centre operators face growing demands on both power and cooling infrastructure. While power resilience has received considerable attention, cooling systems are also evolving in response to changing regulations, rising temperatures and the need for greater energy efficiency.
Balancing resilience with efficiency requires operators to consider how cooling systems respond to sudden changes in demand, while ensuring infrastructure can scale to meet future requirements.
Duncan Singer, Technical Manager – Advanced Products at Excool, discusses the relationship between power and cooling resilience, the impact of AI on cooling requirements and the importance of designing systems for future demands…
Is cooling given enough consideration when the industry talks about power resilience? How important is it to consider power and cooling together when designing a data centre?
Cooling has seen a quieter revolution compared to power when we talk about resilience. F-gas, refrigerant crackdowns and a significant jump in temperatures have changed the face of DC cooling over the past five years, and we still haven’t stabilised. However, heightened temperatures mean increasing energy efficiency through free cooling systems can massively decrease energy consumption.
Where are the biggest vulnerabilities between power and cooling systems? What might operators overlook when planning for power interruptions?
The speed at which cooling capacity needs to be ramped up in case of an AI spike is still under-considered as a risk. To prepare for this, power monitoring at the rack level and having sufficient buffer vessel storage allows for fast reaction to load spikes and inbuilt system resilience.
How can operators balance energy efficiency with resilience? Are there situations where the two can conflict, and how can this be managed?
Luckily, they are mostly synergistic. Part-loading two units allows for lower fan speeds, which significantly reduces energy consumption compared to running one unit at max load. Spreading cooling load over multiple chillers means more of the load can be serviced by free cooling.
How are AI and increasing rack densities changing cooling requirements? Are existing approaches to resilience still fit for purpose?
While we are five years into the AI revolution, the answer is still uncertain, as we are living in a house that is constantly being built and rebuilt. In the current paradigm, higher temps mean more free cooling, so less energy consumption. However, spikes and increases in loads mean the systems have to be bigger. The 80% : 20% load redundancy is now quite firmly replaced with the 90% : 45% load redundancy between liquid and air.
What’s one common misconception about cooling resilience you’d like to challenge?
Resilience at a component level is overrated. You’re adding significant additional CapEx, maintenance time and expenditure on servicing redundant components within cooling plants instead of handling resilience at block level, which simplifies system design, reduces overall CapEx and decreases maintenance time spent.
Looking ahead, what should data centre operators be thinking about now to ensure their power and cooling infrastructure is ready for future demands?
Adopting cooling systems that allow for future scalability and decoupling cooling systems. Using the same loop for air and direct-to-chip cooling loads creates competing requirements within the cooling system. Air cooling requires cold water temperatures, whereas DTC cooling can operate at much higher temps. As a result, chillers are forced to satisfy the coldest requirement, greatly reducing efficiency and free cooling opportunities, all while increasing operating costs.
