top of page

When AI Loads Change in Seconds: Simulating Liquid Cooling Resilience with Cadence Reality DC

Oct 27
4 min read

Updated: 2 days ago

Data centre liquid cooling overview
Data centre liquid cooling overview

AI compute does not ramp gently

One of the challenges of cooling high-density AI infrastructure is that the IT load and the cooling system operate on very different timescales.


A large AI workload can move from relatively low utilisation to full compute in seconds. The heat generated at the rack follows almost immediately. The mechanical cooling system, however, cannot respond at the same speed. Valves need time to open, sensors and controllers need to react, and additional cooling capacity takes time to reach the point where it is needed.


For a short period, the cooling loop itself has to absorb the difference.


That transient period was the focus of a recent Mansim project for a large direct liquid-cooled data centre. The objective was not simply to check whether the system worked at its steady-state design condition, but to understand what happened during the seconds and minutes between a sudden increase in IT load and the cooling system catching up.

Modelling the cooling system as it actually operates

The assessment was carried out using Cadence Reality DC, allowing the cooling network and its transient thermal behaviour to be examined as an integrated system.


The model represented the principal elements of the technology cooling loop, including the direct liquid-cooled racks, coolant distribution units (CDUs), distribution pipework, control behaviour and thermal buffer vessels.


Rather than assessing a single design point, the study followed the system through changing operating conditions. These included rapid increases in compute load, sustained high-load operation, recurring workload cycles and reduced cooling capacity.


This matters because steady-state calculations can tell us whether equipment has sufficient nominal capacity, but they cannot fully describe what happens during the transition from one operating state to another.


For AI data centres, those transitions can be just as important as the final condition.

The importance of thermal buffering

A central part of the study concerned the role of buffer vessels within the liquid-cooling network.


A buffer vessel adds thermal inertia to the system. When rack heat output rises faster than the CDUs can respond, the stored coolant can absorb part of that short-term imbalance.


But volume alone does not determine how well a vessel performs.


The way coolant moves and mixes inside the vessel affects whether it simply delays a temperature rise or actively reduces its severity.


To investigate this, Mansim combined the overall cooling-loop simulation with detailed three-dimensional modelling of different vessel arrangements. This allowed the internal flow patterns, mixing and temperature distribution to be examined rather than treating the vessel as a simple volume of water.


The analysis showed a clear distinction between storage and mixing.


A vessel with limited mixing can delay the arrival of warmer coolant downstream but may do little to reduce the eventual temperature peak. A vessel that mixes more effectively can smooth the transient, reducing both the peak and the rate at which temperature rises.


That distinction becomes particularly important when cooling sensitive high-density compute equipment.

The effect of buffer position

The project also examined where thermal buffering should sit relative to the CDU.

From the racks' perspective, the difference between the configurations was small. But from the cooling plant's perspective, the behaviour was quite different.


When the buffer vessel was positioned in the return path, the sudden temperature step returning from the racks was converted into a more gradual rise before it reached the CDU.


During recurring workload cycles, this reduced the temperature variation experienced by the CDU substantially.


In practical terms, the buffer was doing more than storing thermal energy. It was giving the cooling plant more time to respond.


That is an important distinction in resilient cooling design. The objective is not always simply to minimise the highest temperature. Sometimes the better design is the one that controls how quickly conditions change.

Cooling loop behavior

The Reality DC model was also used to examine how the cooling loop behaved when operating away from its preferred configuration.


One scenario considered fewer CDUs in service. The system continued to maintain acceptable coolant supply temperatures, but the thermal excursion became larger and recovery took longer.


This type of analysis provides a more useful picture of resilience than a simple capacity check.


A system may remain within its absolute temperature limits during a fault or maintenance condition, while still experiencing longer recovery times, higher return temperatures or more demanding conditions at the cooling equipment.

Understanding those differences helps engineering teams define operating strategies and identify where the real margins exist.

From design assumptions to operational behaviour

The value of transient modelling is that it connects several parts of the cooling system that are often considered separately.


Rack heat output, coolant volume, pipework, CDU response, valve behaviour, thermal storage and equipment redundancy all influence one another.

Cadence Reality DC allowed those interactions to be explored before the facility was operating under real AI workloads.


For this project, the overall assessment provided confidence that the liquid-cooling system remained within its critical temperature limits across the scenarios studied, while also identifying how buffer-vessel design, positioning and control strategy influenced the response of the system.


For increasingly dense AI infrastructure, this type of analysis is becoming an important part of cooling-system design.


The question is can it respond quickly enough when the compute load changes?

Comments


bottom of page