AI has changed the behaviour of the load, not just its size. The cooling has to be shown responding when the load changes, as well as carrying it.

A liquid-cooled AI hall is tested in two ranges. Base build testing puts a dependable, controllable heat load into the chillers, CDU and risers while the hall is built, to prove they have the capacity. End-of-line testing replicates the true rack-level heat and electrical load at each rack position, to show how the whole cooling and power system responds, in both directions, before the GPUs are installed.
New loops carry debris that ends up in the cold plates. Clean the loop and prove it clean first.
Confirm flow and pressure at the racks, not just at the CDU.
Heat and electrical load applied together, to prove capacity.
Replicate the racks at each rack position, with the load rising and falling.
It is not enough to put a large amount of heat into a room and say the cooling system has been tested. The system is proven by loading the whole chain and watching how it responds.

Direct-to-chip cooling and AI load arrived together, and each makes the other harder to prove.
The load moves between near idle and full draw quickly, so the cooling has to follow the peaks as well as the average.
The system has to respond when the load suddenly reduces, as well as when it rises.
The electrical and thermal systems are now genuinely coupled, so proving them separately proves neither.
Cold plates and narrow passages are unforgiving about coolant quality.
Not every component in a server is liquid cooled, so some heat still needs air cooling.
Performance varies from rack to rack, so a hall average can hide a rack that does not work.
Proving the riser proves the feed. End-of-line testing replicates the IT electrical and heat load at the rack, which also proves the busway, the tap-off, the local protection and the manifold branch.
Cleaning the loop and proving coolant quality before load.
Read moreLB500, LB600 and LB700 liquid-cooled load banks, plus the ACLB 200.
Read moreLCLB 160, LCLB 240 and ACLB 60, replicating the AI rack.
Read moreDistributes coolant to the load banks.
Read moreUp to 250 units driven and logged from one point.
Read moreRemote support on every project, and on-site support through accredited partners.
Read more


Send the IT load, the cooling design and the programme, and we will tell you what testing it takes.