Cold-plate cooling moves coolant through plates coupled to chips (indirect); immersion contacts devices directly with media. Cold plate suits mainstream AI/HPC (mature, serviceable, cost-controlled — 550 W-class chips measured at 91 °C); immersion suits extreme density but costs more in media and operations.

A CDU is the interface and control node of a liquid-cooling system: a plate heat exchanger thermally isolates the IT secondary loop from the facility primary side, pumps drive coolant to capture chip heat, and filtration, venting, make-up, leak detection and interlocks are integrated. Size it from the loop and fault domain it serves — rack-level L2L CDUs cover 30-500 kW, L2A 10-80 kW.

Pressure drop decides pump sizing and energy: too high needs bigger pumps (more power and noise); too low starves flow and chips overheat. Rack-level systems target ≤80 kPa; the 550 W cold plate measures only 5.8 kPa; manifold branches, fittings and tubing all add pressure drop.

Coolant and wetted materials (metals, seals, hoses, fittings) decide system life: leachables foul microchannels and corrosion products block flow. Run leaching, permeation and aging tests before selection; common media are ethylene/propylene glycol solutions and pure water; seals are EPDM or FKM per media.

Release validation has four levels: component (cold-plate thermal, quick-connect cycling and leak), loop (measured flow/pressure/temperature), rack (full, partial and single-branch maintenance states) and system (interlocks, power recovery, redundancy switchover). Key figures are measured — the 550 W cold plate at 91 °C / 5.8 kPa.

Liquid cooling cuts cooling energy sharply: fully liquid-cooled cooling-side pPUE can reach 1.04 and 50/50 mixed about 1.16 (Inspur public data); marketing figures like "up to 40% power and 40% water savings at 45 °C without chillers" are vendor estimates for specific conditions. Actual PUE depends on load, climate and facility.

An L2A CDU rejects secondary-loop heat to room air or outdoors through a liquid-to-air heat exchanger — deployable without facility water, ideal for retrofits and no-chiller sites. Mainstream units run 70 kW class (10-80 kW); verify room sensible capacity, airflow and noise.

Direct liquid cooling usually covers only major heat sources; memory, power, networking and auxiliary devices still generate heat. Quantify residual air heat and make the boundary explicit: uncovered heat is removed by rack airflow or captured by rear-door heat exchangers (RDHx, 120 kW class), with full-load measurements confirming no hot spots.

Quick connects concentrate leak risk. Per OCP UQD v1.0: working pressure 0-600 kPa, cycle life ≥5,000, fluid loss ≤0.07 mL per cycle (UQD08 limit), 100% helium leak testing at production. Self-sealing on disconnect, bubble-free on connect; blind-mate versions absorb misalignment.

In L2L systems, primary (facility) and secondary (IT loop) water quality must be managed separately: facility water quality, particulates and pressure swings must not enter the secondary. The CDU isolates via a plate heat exchanger; the secondary runs its own filtration (50 μm, optional 25 μm), make-up and monitoring to protect microchannels from fouling and corrosion.

Redundancy and fault domains define availability: N+1 pumps keep the rack running through a single failure (flow/pressure recovery ≤30 s per industry practice), and branch-level isolation lets one node be serviced without stopping the rack. Design makes explicit what fails, what is affected and how recovery works.

Condensation is a hidden risk: when cold-plate surface temperature falls below the room dew point, moisture condenses and can short electronics. Controls: supply temperature above dew point (typically 3-5 °C margin), insulation on cold plates and lines, humidity monitoring and condensation interlocks.

Acceptance runs on measured data across five areas: venting and flushing, water quality, flow/pressure/temperature, leak and power recovery. A 550 W-class cold plate is accepted on measured temperature and pressure drop (91 °C / 5.8 kPa); racks close with full-load three-state validation.