When we opened hall B in ams-1 we chose to cool it with water drawn from a deep lake nearby instead of chillers. A year later we have enough data to say how it went.
The numbers
- Cooling now uses about 9% of the hall's power, down from 31% in our chiller-cooled hall
- GPU temperatures under full load sit 6 degrees lower on average
- Fan and pump failures fell by about half
Summer was the real test
Lake water at depth barely changes temperature through the year. During the July heatwave the intake rose by less than one degree, and we did not have to slow any customer jobs.
The best cooling system is the one you stop thinking about.
What we'd do differently
We underestimated how much filtering the intake would need in spring, when the lake turns over. We now change filters every two weeks from March to May, and plan maintenance windows around it.
Hall C, opening next year, uses the same design with a second intake, so filters can be changed without lowering capacity.
