Introduction
Cold plates keep getting better: finer channels, thinner bases, flow aimed right at the hot spots. Yet plenty of engineering teams mount a new sample, power it up, and find the chip running hotter than the numbers promised. Often the culprit isn’t the cold plate. It’s the thin layer between the cold plate and the chip.
Choosing a thermal interface material for cold plates matters more as cold plates improve. The lower the cold plate’s own thermal resistance, the bigger the share the thermal interface material (TIM) takes of the total. Change the TIM, change the mounting pressure, or let months of thermal cycling slowly pump paste out of the interface, and the same chip on the same cold plate can run several degrees hotter. On an AI accelerator at or above 1 kW, that can be the difference between full performance and throttling.
This guide compares the four TIMs most often used with cold plates (thermal paste, thermal pads, phase-change materials and liquid metal) from a cold plate manufacturer’s point of view: how each works, where it fits, what can go wrong, and what the cold plate side does to make it succeed or fail.
What Is a Thermal Interface Material for Cold Plates?
A thermal interface material for cold plates is a thin layer placed between a chip (or its package lid) and the cold plate base. It fills microscopic gaps between the two metal surfaces and pushes out air, which conducts heat poorly. The four most common types are thermal paste (grease), thermal pads and gap fillers, phase-change materials, and liquid metal.
| TIM Type | Typical Bond Line | Dependence on Mounting Pressure | Main Reliability Risk | Electrically Conductive? | Best Cold Plate Fit |
|---|---|---|---|---|---|
| Thermal paste / grease | Very thin | Medium | Pump-out, dry-out | Usually no (some metal-filled grades are) | Lidded CPUs, general-purpose TIM2 |
| Thermal pads / gap fillers | Thick | High (must be compressed) | Compression set, excess clamping force | Usually no; can insulate | Peripheral components, uneven surfaces, battery modules |
| Phase-change material (PCM) | Thin | Low to medium | Poor contact before first heat-up | Usually no | Server CPUs and GPUs, long-life TIM2 |
| Liquid metal | Extremely thin | Low | Attacks aluminum, diffuses into copper, spills cause shorts | Yes | Bare die (TIM1.5), extreme heat flux |
This is a qualitative comparison. Thermal impedance varies widely by product, pressure and thickness, so vendor numbers shouldn’t be compared side by side.
Where the TIM Sits Between a Chip and a Cold Plate
The TIM that touches a cold plate falls into one of two cases, depending on the package:
- Lidded packages (most server CPUs, some GPUs): Inside the package, the die connects to an integrated heat spreader (IHS) through TIM1. The cold plate then presses onto the lid through TIM2. Compared with TIM1, TIM2 has to cope with far more variation in gap size, tolerance, roughness and contact pressure.
- Bare-die packages (many AI GPUs and accelerators): The cold plate sits directly on the exposed silicon with a single TIM layer, known as TIM1.5, the interface between a lidless bare die and a heat sink or cold plate. It’s common in high-performance computing.
How much of the thermal budget this layer consumes, and how to work back from junction temperature to a cold plate target, is covered in our guide to cold plate thermal resistance. This guide answers a different question: what should that layer be made of?
Why Bulk Thermal Conductivity Is the Wrong Number to Compare
Engineers often shortlist TIMs by thermal conductivity (W/m·K). But the temperature rise across the interface is set by thermal impedance:
Thermal impedance ≈ bond line thickness (BLT) ÷ thermal conductivity + contact resistance on both sides
In other words, a TIM’s total resistance is its bulk resistance (thickness divided by conductivity) plus the contact resistance at each surface. Three practical conclusions follow:
- Thickness often matters more than conductivity. A high-conductivity material that can’t be squeezed thin may lose to an ordinary paste that spreads into a very thin layer.
- Wetting drives contact resistance. A study using the ASTM D5470 method and a custom thermal test vehicle found that a liquid TIM beat a gap pad on thermal impedance at the same bond line thickness, and at every thickness tested, because liquid fills the interface more completely.
- Pressure changes the result. The same grease can post different thermal impedance at 2.75 bar and 6.89 bar, with higher pressure giving lower impedance.
Compare TIMs on thermal impedance at the same pressure and thickness, or on the same test vehicle. A datasheet that lists conductivity without test pressure or thickness tells you very little.
The Four TIM Types Compared
Thermal Paste and Grease
Thermal paste is a silicone or non-silicone base loaded with conductive filler. It flows under assembly pressure into a very thin bond line and is the most common TIM2 on lidded CPU cold plates.
Strengths: Very thin bond lines, low contact resistance, low cost, and easy automation through stencil printing or dispensing.
Risks: The main one is pump-out. As the chip and cold plate heat and cool, they expand at different rates and repeatedly squeeze the interface, pushing paste toward the edges and leaving voids in the center. When the National Renewable Energy Laboratory studied grease reliability, it ran tests designed around pump-out and dry-out conditions and saw voiding, separation and migration. AI servers with long full-load runs and frequent power swings deserve particular attention here.
Best fit: Lidded CPU cold plates with good flatness, stable mounting pressure, and a need for low cost and automated assembly.
Thermal Pads and Gap Fillers
Thermal pads are preformed compliant sheets, usually silicone-based; gap fillers can also be dispensed as a gel that cures in place. Both are thick and rely on compression to conform.
Strengths: They absorb large height differences and tolerances, assemble cleanly, can be reworked, and most provide electrical insulation.
Risks: Thickness alone usually makes their thermal impedance noticeably higher than paste. Pads also need enough compression to make full contact, and that compression pushes back on the components and the cold plate. Under long-term load, they can take a compression set and lose contact.
Best fit: VRMs, inductors, driver ICs and other uneven components around the main chip; battery modules, where insulation also matters.
Phase-Change Materials (PCM)
Phase-change materials are solid sheets or pastes at room temperature. Once the chip heats up, they soften and flow into the interface, then turn solid again as they cool.
Strengths: They install as cleanly as a pad, then thin out like a paste. According to vendor data, they reach a thin bond line with relatively low pressure and resist pump-out under thermal cycling, which makes them a common TIM2 choice for server CPUs and GPUs.
Risks: Until the first heat-up, the material hasn’t fully flowed and contact resistance is higher, so plan for that in factory testing and first power-on. Softening temperatures vary by product; confirm the chip runs hot enough to complete the phase change.
Best fit: TIM2 on server CPUs and GPUs where long-term stability matters.
Liquid Metal
Liquid metal TIMs are typically gallium-based alloys that are liquid at or near room temperature. They conduct heat far better than polymer-based materials, wet almost any surface, and form extremely thin bond lines.
Strengths: The lowest thermal impedance of the four, a strong fit for bare dies with extreme heat flux.
Risks: It conducts electricity, it flows, and it reacts with some metals, so it can’t simply replace paste (see the next section).
Best fit: Mainly bare dies (TIM1.5) and in-package interfaces. Even a liquid metal supplier describes its products as typically used for TIM0 and TIM1 applications, the interfaces closest to the die. If a lidded package needs a metal TIM2, compressible solid indium is another option; because it stays in a solid metal state, it helps mitigate pump-out and bake-out.
Liquid Metal on a Cold Plate: Compatibility Comes First
Liquid metal puts stricter demands on the cold plate than any other TIM. Before choosing it, answer four questions:
1. Is the contact surface aluminum? Nearly every common liquid metal alloy contains gallium, which is extremely aggressive toward aluminum. In one test described on Indium’s blog, a gallium-indium alloy was applied to an aluminum heat sink and held at 125 °C for seven days; the gallium dissolved through the aluminum grain boundaries and destroyed the heat sink’s integrity. Liquid metal should never touch bare aluminum.
2. Does a copper base need plating? Copper holds up far better, but gallium diffuses into copper, slowly consuming liquid metal at the interface. The usual answer is a nickel-plated contact area, since gallium is non-reactive with nickel.
3. How will spills be contained? Liquid metal is conductive and can short nearby circuitry. Research on liquid metal pastes notes that a liquid metal TIM has to be contained to prevent leaks, and nearby components should be insulated. That shapes the contact-area outline and whether a gasket or dam is needed.
4. What about rework? Cleaning and reapplying liquid metal is far messier than paste, so plan it into service procedures.
Choosing liquid metal isn’t a material swap. It’s a system decision that touches cold plate material, surface finish, contact-area geometry and maintenance.
Mixing TIMs on One Cold Plate: Main Package vs. Peripheral Components
Many server cold plates also cool the VRMs, inductors or switch chips around the main chip, so one cold plate carries two kinds of TIM:
- Main package area: A thin paste, PCM or liquid metal for the lowest thermal resistance. On an AI GPU, this area usually covers the GPU die and the HBM stacks in the same package under one shared TIM layer.
- Peripheral areas: Thermal pads or gap fillers to absorb height differences and tolerances.
The most common failure is pad compression force. If a pad is too thick or the pedestal above it is off, the pad’s pushback lifts the cold plate. The TIM over the main package gets thicker, and the most critical chip runs hotter.
That’s why cold plate bases are often stepped or built with pedestals, each contact zone set to its component’s height and tolerance so pads see only a reasonable compression and the main package keeps full contact. That takes component height tolerances plus the thickness and compression behavior of every TIM involved. For deciding what a GPU cold plate should cover, see our server GPU cold plate selection guide.
Which TIM for Which Cold Plate Application?
| Application | Common TIM Choices | Key Considerations |
|---|---|---|
| AI GPUs and accelerators (bare die) | PCM or high-performance paste; liquid metal for extreme heat flux | Concentrated hot spots; liquid metal needs a plated contact area and spill containment |
| Server CPUs (lidded) | Paste, PCM | Mounting pressure is capped by the socket spec; watch for pump-out under sustained load. See our server CPU cold plate selection guide |
| IGBT and SiC power modules | Thermal grease, PCM | Follow the module maker’s mounting instructions; directly cooled pin-fin modules need no TIM. See our IGBT cold plate design guide |
| EV and energy storage battery cold plates | Thermal pads, gap fillers | Large areas, loose tolerances and electrical insulation. See our battery liquid cold plate guide |
| Laser and optical modules | Paste, PCM, indium foil | Tight temperature stability; avoid outgassing onto optics |
What the Cold Plate Maker Controls
Half of a TIM’s performance comes from the material. The other half comes from the two surfaces around it and how hard they’re pressed together. On the cold plate side, these factors matter most, and APALTEK builds each of them into the design and quote to match customer requirements:
- Contact-area flatness and roughness. An uneven surface forces the TIM to get thicker. APALTEK holds the contact area to the flatness and roughness on your drawing or your chip vendor’s spec and confirms them during DFM review.
- Contact surface finish. If your TIM or coolant calls for a nickel-plated contact surface, as liquid metal does, specify it in the RFQ and we’ll build it into the process route and quote.
- Pedestals and stepped bases. Pedestal height and footprint are designed around your package outline, peripheral component heights and planned TIM bond line.
- Mounting pressure and retention. Pressure determines how thin the TIM gets. It must stay within the chip vendor’s package loading spec, and the cold plate’s stiffness and fasteners determine whether it stays even over the product’s life.
- The TIM used in testing. When thermal resistance is defined to include the TIM (case-to-inlet, for example), the result depends on which TIM and pressure were used. For samples, we can test with the TIM and mounting load you specify, so the data reflects your real interface. Test scope follows your requirements; see our testing and validation capabilities.
How to Specify the TIM in Your Cold Plate RFQ
If your thermal targets depend on the TIM, include at least:
- Package type: lidded (TIM2) or bare die (TIM1.5)
- The TIM type and specific product you plan to use
- The chip vendor’s mounting load or pressure range
- Contact-area flatness, roughness and surface finish
- For peripheral components: heights, tolerances and planned pad thicknesses
- Whether thermal testing should use a specified TIM
For the full list, see our custom cold plate RFQ checklist.
FAQ
Does a cold plate always need a TIM?
Almost always. Even a well-machined cold plate touches the chip only at microscopic high points, with air in between. The TIM fills those gaps. The main exception is a power module with a pin-fin baseplate cooled directly by coolant, where there’s no interface to fill.
Can liquid metal be used on an aluminum cold plate?
Not directly, because gallium attacks aluminum. Use a copper cold plate with a nickel-plated contact area, or a reliable barrier layer if aluminum is unavoidable.
Can thermal pads replace thermal paste?
Not on the main chip, where a pad’s extra thickness costs real thermal performance. Pads are the better choice on peripheral components or wherever height differences are large.
How often does a TIM need to be replaced?
It depends on the material, operating temperature and thermal cycling. The industry still lacks a unified method for characterizing TIM degradation in data center platforms, so run thermal cycling and aging tests during validation and track chip temperature trends in the field.
Does a cold plate supplier’s thermal resistance figure include the TIM?
It depends on the definition. Case-to-inlet figures include the TIM; surface-to-fluid figures don’t. Before comparing suppliers, confirm the definition, TIM and mounting pressure all match.
Conclusion
No single TIM is best. Paste, pads, phase-change materials and liquid metal each have their place, and interface resistance comes from the combination of material, bond line thickness, mounting pressure and the cold plate surface. Liquid metal in particular puts its own demands on cold plate material and finish.
If you’re developing cold plates for AI servers, power electronics or other high-power equipment, send APALTEK’s engineering team your package type (lidded or bare die), power and temperature limits, planned TIM, mounting load, and whether the cold plate must cover peripheral components. We’ll evaluate contact-area flatness, surface finish and pedestal design against those conditions, and for samples, we can measure thermal resistance with your specified TIM and report the test conditions alongside the data.