A direct-to-chip liquid-cooling architecture for high-power AI processors, combining an aluminum cold plate, optimized internal flow channels, and a modular rack architecture—designed not just to cool the machine, but to explore how efficiently we can move heat before the silicon hits its limits.
- The problem: AI accelerators are actively pushing towards power intensive procedures. The challenge is no longer simply removing heat. Nope. it is moving enormous amounts of heat uniformly and efficiently without creating localized hotspots, excessive pressure drop, or unsustainable pumping and cooling loads.
- The solution (potentially): A liquid-cooling system that optimizes itself. Done by using real-time thermal feedback and trained AI to continuously reshape flow distribution across parallel microchannels. Flow patterns would be determined by current condition of the chip for example identified hotspots or pressure drops.
- The progress: Simulation of a data centre condition has been done with aluminum extrusion racks. A small 1.5hp pump for water circulation installed and a liquid cooling radiator fan to dump excess heat. Potential micro channels to effectively remove heat are under testing.
SolidWorksANSYSMATLABHeat Transfer
- Target heat flux
- ~100 W/cm²
- Plate material
- Al 6061-T6 (testing)
- Working fluid
- Water
- Measurement
- Thermocouples
Drawings & photos · 04
CADAI Generated model of a potential cold plate flow channel
PHOTOModel data centre rack with aluminum extrusions
CADInstalled radiator dimensions
VIDEOSolidworks assembly
Digital renderings · 00