Open a modern AI server tray and the first thing you notice is not a heatsink fin stack but a flat metal block bolted to the GPU package, with two small hoses leaving it through colored quick disconnects. That block is a cold plate, and the fluid moving through it is the front end of a direct-to-chip liquid cooling system.
Direct-to-chip liquid cooling, often shortened to DLC or D2C, is a method that pumps a single-phase or two-phase coolant through sealed cold plates pressed against high-power silicon, carrying heat out of the server before it ever reaches the room air. The rest of this article follows that coolant from the die surface to the facility chiller, then works backward through the design and retrofit decisions that determine whether the loop actually performs.
Основные выводы
- The defining feature is contact: a cold plate sits on the CPU, GPU, or accelerator, not on a fin stack cooled by fans.
- A coolant distribution unit (CDU) isolates the server-side fluid from facility water through a heat exchanger, not just a manifold.
- Roughly 20–30% of server heat typically still needs air handling, so DLC supplements room cooling rather than replacing it.
- Two-phase variants exploit boiling for higher flux density; single-phase variants are simpler to commission and service.
- Retrofit feasibility hinges on three unsexy items: floor loading, secondary loop water chemistry, and rack-level leak detection.
Direct-to-Chip Cooling in One Sentence
Direct-to-chip liquid cooling moves heat off the hottest packages in a server by clamping a liquid-filled cold plate onto the chip lid and routing that liquid, through quick disconnects and a rack manifold, into a CDU that hands the heat to facility water. Everything else — fans, rear-door heat exchangers, CRAH units — exists to handle whatever heat the cold plates do not capture.
Follow the Heat From GPU to Facility Loop
Start at the silicon. A 700 W accelerator dumps almost all of its electrical input as heat into a thin layer of solder and copper under the package lid. The cold plate sits on top of that lid with a thin film of thermal interface material between them, and inside the plate a pattern of microchannels — typically 200–500 µm wide — forces coolant into turbulent contact with the copper base.
Heat conducts from die to lid to plate to fluid in that order, and any one of those interfaces can dominate thermal resistance if it is built poorly.
From the cold plate, two hoses leave the chip — supply and return — and join a tray-level manifold. The tray manifold ties into a vertical rack manifold, sometimes called a CDM, through dripless quick disconnects that allow servers to be pulled without draining the loop. CoolIT and similar vendors publish cold plate specifications showing pressure drop and thermal resistance per node, which the rack designer uses to size flow (CoolIT Systems cold plates).
The rack manifold then feeds the secondary loop of a CDU. Inside the CDU, a brazed-plate or shell-and-tube heat exchanger transfers heat from the technology cooling system (TCS) loop to the facility water system (FWS) loop. The facility loop carries that heat to a dry cooler, cooling tower, or chiller plant — the same equipment that already serves CRAHs, just sized for higher return temperatures.
Vertiv’s overview of liquid cooling architectures describes this two-loop separation as standard practice (Vertiv liquid cooling options).
What the Cold Plate Changes Inside the Server
Replacing a heatsink with a cold plate changes the entire thermal layout of the tray. Fans shrink or disappear above the cooled packages, freeing static pressure budget. Memory, VRMs, NICs, and power supplies, however, still need airflow — direct-to-chip does not touch them by default.
Cold Plate Geometry and TIM Choice
Microchannel density determines how much surface area the coolant sees. Skived-fin and 3D-printed plates push that density higher than older pin-fin designs, which matters as die-level power density climbs. The thermal interface material under the plate is a recurring failure mode: pump-out and dry-out of grease-based TIMs under thermal cycling is a well-known mistake on early DLC retrofits, and many operators have moved to phase-change pads or indium foils for accelerators above roughly 500 W.
Residual Air Heat
A cold plate captures the package, not the board. Power that leaves the VRMs, DIMMs, optics, and storage still becomes air heat inside the chassis. Most current direct-to-chip deployments still rely on chassis fans plus room cooling — or a rear-door heat exchanger — to remove that residual fraction.
Sizing the air path for the remaining heat, not the original heat, is where retrofit calculations usually go wrong.
Why a CDU Is More Than a Pump Box
The CDU does four jobs at once: it isolates the clean secondary loop from facility water, regulates supply temperature, controls flow and pressure to the racks, and provides the instrumentation needed to detect leaks and degradation. Treating it as "just a pump skid" is the procurement mistake that lands operators with undersized heat exchangers six months into a deployment.
CDU Function | What It Actually Does | What Breaks If You Skip It |
|---|---|---|
Теплообменник | Hands heat from TCS to FWS loop | Facility water contamination from glycol or biocide |
Variable-speed pumps | Match flow to IT load | Hot spots at high-power racks, wasted pump energy at low load |
Supply temperature control | Mixes return to hit setpoint | Condensation on cold plates when room dew point is high |
Filtration and side-stream treatment | Holds particulate and biological growth in check | Microchannel fouling, gradual flow loss per node |
Leak detection and isolation valves | Detects pressure or conductivity changes | Coolant reaching a powered busbar |
Sizing follows the rack power, the approach temperature of the heat exchanger, and the available facility water temperature. Higher facility water temperatures — sometimes 30–40 °C supply — are what make DLC interesting economically, because they unlock dry-cooler operation for much of the year. The Uptime Institute’s 2024 cooling survey documents the shift toward warmer secondary loops as direct liquid cooling adoption grows (Uptime Institute cooling systems survey 2024).
What Direct-to-Chip Does Not Cool
This is the question that decides whether a deployment succeeds. Direct-to-chip handles the CPU, GPU, and sometimes the HBM stack or switch ASIC — anything a cold plate can be mechanically mated to. It does not handle:
- DIMM banks beyond what a secondary cold plate or air can reach
- Optical transceivers in the front of the chassis
- Power supplies and their internal magnetics
- Storage drives, NICs, and BMC components
- The PDUs, busbars, and rack-level power gear
The practical consequence is that a "liquid-cooled" rack is rarely 100% liquid-cooled. A realistic capture ratio for current DLC servers is in the 70–80% range, with the balance handled by chassis airflow. Immersion cooling is the alternative when capture ratios approach 100%, but it is a different architecture with different fluid, materials, and serviceability tradeoffs.
Single-Phase vs. Two-Phase Direct-to-Chip Designs
Both variants use cold plates, manifolds, and a CDU. The difference is whether the working fluid stays liquid or changes phase inside the cold plate.
Атрибут | Single-Phase DLC | Two-Phase DLC |
|---|---|---|
Working fluid | Water with corrosion inhibitor, or PG/water blend | Engineered dielectric with low boiling point |
Heat transfer mechanism | Sensible heating of the liquid | Latent heat of vaporization at the chip |
Pressure regime | Pumped, modest pressure | Often near-atmospheric, condenser-driven |
Flux capability per cold plate | Bounded by flow rate and ΔT | Higher, since boiling moves more heat per gram |
Fluid cost and handling | Lower, well-understood chemistry | Higher, regulatory attention on PFAS-class fluids |
Field service complexity | Familiar to most operators | Requires condenser design and phase-aware controls |
Single-phase dominates current shipments because the chemistry, materials, and service procedures look like building HVAC work. Two-phase is the path forward when chip power densities outrun what a pumped liquid loop can carry, but the engineering and supply chain are less mature, and the fluid landscape is still in flux.
Retrofit Questions Before the First Rack Arrives
Greenfield DLC deployments can pick their floor, piping, and water chemistry. Retrofits cannot, and that is where most pain originates. Walk these questions before signing a rack order.
Can the Floor and Slab Take the Weight?
A populated DLC rack with onboard CDU support can exceed 1,500–2,000 kg. Raised floors built for 800 kg per tile will not survive without structural reinforcement, and the load path matters as much as the total weight.
Where Does the Facility Water Come From?
If the building already has chilled water at 7 °C, you are paying chiller energy to feed a loop that would happily run at 32 °C. A tie-in to a warmer condenser water loop, or a dedicated dry-cooler circuit, is usually a better answer — and it changes the CDU heat exchanger sizing.
What Is in the Secondary Loop Water?
Microchannels are unforgiving. Water quality requirements typically specify conductivity, hardness, dissolved oxygen, and biocide chemistry compatible with copper, nickel plating, and EPDM hoses. A common installation mistake is filling the secondary loop with treated building water and skipping the flush-and-passivate step, which seeds the loop with debris that lodges in the smallest channels first.
How Will a Leak Be Detected and Contained?
Drip trays, conductive rope sensors at manifold joints, and CDU pressure-decay logic are the standard layered approach. Rack-level isolation valves that close on detection are worth the cost; relying on a human to hit a button is not.
Вопросы и ответы
Is direct-to-chip cooling the same as immersion cooling?
No. Direct-to-chip uses cold plates and sealed loops, so only the chip surface meets coolant indirectly through metal. Immersion submerges the entire board in a dielectric fluid.
The two can coexist in the same data hall but are not interchangeable retrofits.
Can I run direct-to-chip on warm water without a chiller?
Often yes, depending on climate and chip thermal limits. Many current GPUs accept coolant inlet temperatures in the 30–40 °C range, which lets dry coolers handle the facility side for much of the year. The exact ceiling is set by the silicon vendor’s junction temperature limit and the cold plate’s thermal resistance, not by the cooling architecture itself.
What happens to my existing CRAC units after a DLC retrofit?
They keep working, just lighter loaded. Residual air heat from memory, optics, and power supplies still has to go somewhere, and rooms with mixed liquid-cooled and air-cooled racks need their CRACs balanced for the new airflow pattern. Operators who decommission CRACs too early end up with hotspots on adjacent legacy rows.
How often does the secondary loop need maintenance?
Filter changes, biocide checks, and conductivity sampling are typically scheduled quarterly, with a deeper inspection of the CDU heat exchanger annually. The interval depends on fluid chemistry and run hours, and CDU vendors publish the procedures specific to their fluids. Skipping side-stream filtration is the cheapest way to ruin a year-old loop.
Do quick disconnects really not drip?
Modern dripless QDs on DLC servers lose a few drops at most during disconnect, and that volume is designed to evaporate without damaging hardware. They are not the same parts as garden-hose couplers, and substituting cheaper fittings during a service event is a documented cause of in-rack contamination.
Заключение
If someone points at a rack of AI servers and asks what is direct-to-chip liquid cooling, the honest answer starts at the cold plate, walks through quick disconnects and a rack manifold into a CDU, and ends at facility water leaving the building through a dry cooler or chiller. Everything technical about the architecture — single-phase versus two-phase, water chemistry, residual air heat, CDU sizing — is downstream of that one heat path.
Get the path right, plan for the heat the cold plates do not capture, and the rest of a direct-to-chip deployment is engineering rather than guesswork.