Modern AI accelerators and tightly coupled server nodes are pushing power far beyond what traditional air cooling can handle. Individual chips can exceed 1,000 W, and a single rack can dump more than 100 kW of heat into the room. No amount of fans or clever ducting can remove that efficiently with air alone.
This is where single-phase direct liquid cooling (DLC) comes in. The idea is straightforward: circulate water or a water–glycol coolant through coldplates mounted directly on the hottest components – CPUs, GPUs, and AI accelerator modules. The coolant absorbs heat at the coldplate, then carries it away in a closed loop to a coolant distribution unit (CDU), where the heat is rejected to facility water or another secondary loop.
Why single-phase DLC works so well
Compared to air, liquids:
- store far more heat per unit volume,
- move heat away much faster,
- allow tighter temperature control at higher power levels.
That translates into:
- higher chip TDP per socket,
- greater rack density in the same footprint,
- more compute per square meter of data center space.
Versus two-phase and immersion cooling
The paper also contrasts single-phase DLC with two-phase (boiling, phase‑change) and immersion cooling (submerging entire boards in dielectric fluid). While those approaches can deliver even higher heat flux removal, they introduce more complex infrastructure, specialized fluids, and tougher serviceability.
Single-phase direct liquid cooling strikes a balance: it uses relatively simple water-based coolants, integrates with existing rack and server designs, and is already proven for the next decade of rising processor power and rack density in high‑performance and AI data centers.
Source: IEEE Spectrum
