Perfiles de aluminio de placa de refrigeración líquida para servidores de IA

Respuesta breve: an AI server cold plate is a multi-channel aluminum extrusion that carries coolant directly under the chip, with the flow channels formed in the extrusion rather than welded in afterwards. Single-chip TDP has moved from 700 W to over 1,000 W, and at that power the cold plate stops being a heatsink and becomes a pressure vessel that happens to conduct heat. We hold base flatness at ≤ 0.03 mm/100 mm, channel width from 3 mm, wall thickness within ± 0.1 mm, and a helium leak rate below 1×10⁻⁶ Pa·m³/s. Tolerance and channel geometry are set against your drawing and confirmed before the die is cut.

Why the industry moved from welded plates to one-piece extrusion

A welded cold plate is made by joining a machined channel plate to a cover plate. It works, and it was the standard for years. What changes the calculation is the duty cycle: a server cold plate sees continuous thermal cycling, pump vibration and pressure pulsation for years, and every weld is a fatigue initiation site. The failure mode is not a burst; it is a slow weep at a seam that appears after thousands of cycles, inside a rack, above equipment that cannot get wet.

One-piece extrusion removes the seam instead of inspecting it better. The channels are formed by the die, so the coolant path has no joined boundary anywhere along its length. There is no weld to fatigue, no heat-affected zone with different metallurgy, and no flux residue to corrode from the inside. That is the structural argument, and it is the reason the process replaced welded construction rather than merely competing with it.

The thermal argument follows from the geometry. A welded plate needs a flat mating interface between the two halves, and that interface adds contact resistance between the coolant and the base. An extruded channel has the wall and the base as one continuous piece of metal, so heat travels from the die surface to the coolant without crossing a joint. Measured against a welded equivalent, the extruded section shows over 15% better heat exchange efficiency at the same footprint, which at these power levels is the difference between a chip running at spec and one throttling.

Especificaciones

Elemento	Rango
Construction	Multi-channel one-piece extrusion, no welded seams along the flow path
Aleaciones	6061-T6 high-strength, 3003 corrosion-resistant
Ancho de sección	50–600 mm
Channel structures	Serpentine, split-flow (parallel), microchannel
Ancho mínimo de canal	3 mm
Tolerancia dimensional del canal	Reference ±0.05 mm, set by your drawing
Uniformidad del espesor de pared	±0.1 mm
Planitud de base	Reference ≤ 0.03 mm per 100 mm
Tolerancia de referencia de montaje	Reference ±0.03 mm
Alabeo total	≤ 0.05 mm
Cavity machining accuracy	±0.02 mm by wire EDM where required
Precisión de mecanizado CNC	±0.01 to ±0.005 mm
Leak rate	Below 1×10⁻⁶ Pa·m³/s by helium mass spectrometry
Heat exchange gain over welded	Over 15% at equal footprint
Internal corrosion protection	Passivation or anodizing on the channel surfaces
Service life with glycol coolant	Over 8 years
Norma de referencia	GB/T 6892-2023, high-precision grade
Deep processing	Five-axis milling, sealing-face grinding, drilling and tapping, helium leak test

La planitud es un parámetro térmico, no cosmético

At 1,000 W, the interface between the chip package and the cold plate is where a surprising share of the temperature budget is spent. Nobody machines two surfaces flat enough to touch across their whole area; the real contact happens at microscopic high points, and the rest of the gap is filled with thermal interface material. The thickness of that material is set directly by the flatness of the plate.

A plate at 0.03 mm per 100 mm keeps the gap thin and even, so the interface material does the job it was chosen for. A plate that warps to 0.1 mm forces a thicker layer to bridge the gap, and every extra micron of interface material adds thermal resistance that no amount of fin area can recover. On a 1,000 W device, a flatness problem that looks like a machining detail shows up as a junction temperature the cooling loop cannot pull down.

This is also why overall warpage is specified separately from local flatness. Local flatness governs the contact patch under the die; warpage governs whether the plate still seats against the mounting frame after the fittings are torqued. A plate can be flat at every measured point and still lift at one corner once installed, which opens a gap exactly where the hottest die sits.

The channel trade-off nobody gets for free

Narrower channels and thinner walls transfer heat better: more surface area per unit volume, shorter conduction path through the metal. They also raise pressure drop, because hydraulic resistance scales roughly with the inverse of the channel dimension raised to a high power. A pump sized for a serpentine plate with generous channels will not deliver the same flow through a microchannel design, and lower flow means warmer coolant returning to the plate.

So channel width is not a number to minimize. Three millimeters is the practical floor for this class of plate: below it, the gain in heat transfer coefficient is partly given back in pump power and flow reduction, and the channel becomes sensitive to particulate in the loop and to blockage during filling. A serpentine layout gives long residence time and even coverage at the cost of pressure drop; a parallel split-flow layout cuts pressure drop and can suffer maldistribution if the inlet manifold is not designed carefully. We run CFD on the channel topology before the die is cut, because correcting a manifold after the die exists is not a machining fix.

Coolant, corrosion, and the eight-year number

Data center loops run on water-glycol, usually with inhibitors, and the glycol is there for freeze protection and biological control rather than for the aluminum. Aluminum in a wet loop is a corrosion cell waiting for a cathode, and the inside of a cold plate is a set of narrow channels where deposits form first and are hardest to see.

Passivation or anodizing the channel surfaces raises the electrochemical corrosion resistance several times over, and that is what makes an eight-year service life credible instead of optimistic. It also matters that the extrusion is clean internally: flux residue, machining swarf or drawing lubricant left in a channel will start a corrosion pit that no surface treatment can stop once the plate is sealed. This is why the process is sequenced as extrude, machine, treat, clean, leak test, and not as a series of separate purchases.

The helium test is the last gate. A rate below 1×10⁻⁶ Pa·m³/s is roughly three orders of magnitude tighter than what a bubble test can find, and it is the only practical way to certify a plate that will sit above a rack of servers. A plate that passes a pressure decay test and fails helium is a plate that would have wept in service.

Dónde encaja esto

• AI training servers and GPU clusters, where per-chip power has outgrown air cooling and the cold plate is on the critical path
• Direct-to-chip liquid cooling loops in data centers, where a leak is a facility-level event rather than a component failure
• High-density compute chassis where the plate has to be both a thermal component and a structural mounting surface
• Applications where the coolant is water-glycol and the loop cannot be drained and inspected on a short interval
• Designs that need long continuous plates, where a welded assembly would introduce several seams per unit

Dónde no encaja

• Low-power devices under roughly 200 W, where a copper heat pipe or a plain extruded heatsink is cheaper and adequate.
• Geometry that cannot be extruded: a channel that has to change direction within the plane of the plate, or a plenum that must branch in three dimensions. Those still need machining or welding.
• Very small production volumes where the die cost cannot be amortized. A machined plate from billet is more expensive per part but has no tooling.
• Loops using coolants that attack aluminum regardless of coating, such as certain high-chloride or ammonia-based systems.

The honest dividing line is channel complexity. If the flow path is straight and constant along the extrusion direction, extrusion wins on cost, reliability and thermal performance. If the flow path has to turn, the design has to accept a joint somewhere, and then the conversation becomes about how that joint is made rather than whether it exists.

Qué enviar para una cotización

• The thermal requirement: chip TDP, allowed junction temperature, coolant inlet temperature and flow rate
• The mechanical interface: die footprint, mounting hole pattern, and the fitting type and position
• The channel layout if it is already defined, or the heat load and footprint if it is not
• Flatness and warpage requirements for the die contact face, and which face is the datum
• The coolant chemistry, so the internal surface treatment can be chosen for it
• Whether helium leak testing is required per unit, and to what rate
• Quantity, which determines whether a die is justified

If the thermal design is still open, send the chip specification and the rack constraints. Channel topology and section can be proposed from that, and the CFD result will show whether the layout meets the pressure drop budget before tooling is committed.

Preguntas frecuentes

What flatness and channel tolerance can you hold?

As a reference, base flatness ≤ 0.03 mm per 100 mm, overall warpage ≤ 0.05 mm, channel width tolerance ±0.05 mm and wall thickness uniformity ±0.1 mm. Those figures are achievable on 6061-T6 and 3003 sections in the 50–600 mm width range, and they move with section complexity and wall thickness. The achievable values for your specific section are confirmed against your drawing before the die is cut, and if a requirement cannot be met we say so at that stage rather than at first article.

How is a one-piece extruded plate better than a welded one?

It has no seam along the coolant path, so the fatigue and corrosion failure modes that start at a weld do not exist. Thermally, the wall and base are one piece of metal, which removes the contact resistance of the joined interface and gives over 15% better heat exchange at the same footprint. Practically, it also removes a process step and a quality risk: there is no weld inspection to argue about.

Can the flow channels turn a corner?

Not within the extrusion. The channel runs straight along the extrusion direction, and direction changes are made by machining the ends, by a manifold block, or by joining sections. That is the real design constraint of the process, and it is better raised during thermal design than after the layout is fixed. Many layouts can be reworked into parallel straight channels with a machined inlet plenum and lose very little performance.

What coolant can it run?

Water-glycol, typically with inhibitors, is the standard. The channel surfaces can be passivated or anodized to raise corrosion resistance, which is what supports an eight-year life in a closed loop with glycol. Coolants outside that chemistry, particularly high-chloride or ammonia-based fluids, need to be checked against the alloy before the plate is specified.

Is helium leak testing done on every unit?

It can be, and for data center work it usually should be. The test certifies a leak rate below 1×10⁻⁶ Pa·m³/s, which is well beyond what a pressure decay or bubble test resolves. Whether it is per unit or per batch is a commercial decision, but the specification should say which, because the two are not equivalent as evidence.

What is the die lead time?

Tooling for a new section is the long pole in the schedule, and it depends on section complexity, channel count and width. Rather than quote a number that will not hold, we confirm the tooling schedule against the actual section drawing once it is received, and it is stated with the quotation. For an existing section, no tooling is needed at all.

Envíe su plano

Email linjunbing@truetrustlv.com with the section drawing, the chip TDP and the coolant conditions. If the thermal design is still open, send the footprint and the heat load and we will propose a channel topology. You will get a process proposal, a quotation and the tolerances we can actually hold — reviewed by an engineer, not an auto-reply.

WhatsApp: +86 136 6644 0600