The 3D rig needs WebGL, which this browser has turned off.
A test rack cooled by fans and a liquid loop, and a controller that decides every second how much heat each one carries. It picks the split that uses the least power, keeping chips under 70 °C and the coolant above the dew point.
Heat
Cartridge heaters in machined aluminum blocks follow a 45-minute load recording from our own server, watt for watt. Every run sees the same work, so the only thing that changes is how it gets cooled.
The lower two nodes are air-cooled. The upper two carry cold plates, like a hall halfway through a liquid retrofit.
Air
Fan power rises with roughly the cube of fan speed. Slow the fans by a fifth and they draw about half the power. Every watt the liquid loop can safely take lets the fans slow down.
Pfan ∝ N³ 0.8³ = 0.51
Liquid
A Peltier stage chills the coolant below room temperature. Push it below the dew point of the room air and water beads on the cold plates, right above live electronics.
The coldest safe coolant is the dew point plus 2 °C, and it moves with the weather.
Control
Runs offline on a Raspberry Pi 5 and an ESP32-S3. No cloud and no reinforcement learning.
Try it
Dew point, Magnus formula
In a room at °C with % humidity, water condenses below --. So the coolant can't go under --, and the liquid loop carries about -- of the heat.
Try , or . The liquid share is an illustrative curve, not a rig measurement.
What's in the rack
About $455 to $500 in parts
Heat
Four CNC aluminum blocks
Cartridge heaters
Replay recorded server load
Air
PWM fan wall
Finned sinks on the lower nodes
Radiator fans on the roof
Liquid
PWM DDC pump
Aluminum cold plates and radiator
Inhibited coolant
Chiller
Two TEC1-12706 Peltier modules
Pulls coolant below room temperature
Sensing
SHT31 for the dew point
DS18B20 on every node
INA219 power monitors
Control
ESP32-S3 for PWM and sensors
Raspberry Pi 5 for the model
Logs everything at 1 Hz
Safety
KSD301 thermal fuses
Leak sensor on a hardware relay
24-hour dry leak test first
Limits
Chips under 70 °C
Coolant at least 2 °C above the dew point
Progress
Updated Oct 14, 2026
Done
Eight-week rig plan starts
246,934,820 rows of Alibaba cluster data scanned for load samples
First hour on hardware: 3,601 rows at 1 Hz, zero gaps
2,700 s of real server load recorded. It drives the rig above
Next
Cold plate water test and a 24-hour dry leak test
Our controller against a tuned baseline: humidity sweeps, load spikes, condensation safety
Four-rack CFD model, then a fast surrogate for judges to try live
Cooling energy saved at equal safety: measuring now
Questions
The ones judges ask
A thermostat holds one temperature. Delta-T chooses between two cooling systems with different costs, under two limits that move: chip temperature and the dew point.
We don't know yet, and we won't guess. It gets measured on the rig, for the same workload at equal safety, against a properly tuned standard controller.
It would be harder to trust. A fitted physical model and a small search over fan and pump settings is fast, explainable and easy to check against the limits.
Thermal fuses and a leak sensor on a hardware relay cut power without asking the software. The controller is never the last line of defense.
Operators of older air-cooled data centers adding liquid cooling for AI hardware, who have to run both systems in the same room.
A 45-minute recording from our own server, replayed through the heaters watt for watt, so every test run does the same work.