Most explanations of thermal management in high-frequency boards begin with the idea that high frequency generates more heat, and then recommend a laminate with better thermal conductivity or an aluminium substrate. Both halves of that are wrong in a way that costs money. Frequency does not generate heat; the amplifiers, the power conversion and the copper losses do, and on a 10 W RF board the dielectric contributes a fraction of a watt while the output stage contributes fifteen. This explanation works the problem as a heat account in three layers: where the heat comes from, how much resistance it meets on the way out, and which of the available paths actually removes it.
A trace carrying a signal at 10 GHz dissipates the same power as the same trace carrying the same current at 10 MHz, apart from the change in conductor resistance from skin effect and the dielectric loss in the laminate. Both of those are small at signal level. A 50 Ω line carrying 1 W at 10 GHz on a 0.0027 laminate loses about 0.32 dB/in, which is 7 % of the power per inch, or 0.07 W/in. A board with ten such lines and 100 mm of each dissipates under 3 W in total.
The heat on a high-frequency board comes from four places, and only one of them is inside the transmission line:
| Source | Typical share | What sets it | Where it must go |
|---|---|---|---|
| Output stage (PA, driver) | 60–85 % | Output power and drain efficiency: 10 W at 40 % is 15 W of heat | Down through vias into a spreader |
| Power conversion | 5–15 % | Efficiency of the regulator or converter, usually 85–95 % | Copper area and planes |
| Interconnect loss | 2–10 % | dB per inch, line length, power level | Local, into the laminate |
| Digital core and memory | Application dependent | Device dissipation, not the board | Package, then board, then spreader |
The practical consequence is that a thermal design for a high-frequency board is almost always a design for one component — the output stage — and the rest of the board is a heat-spreading and isolation problem. A laminate change is worth evaluating only after that is understood, because it addresses the smallest term.
Heat leaving a die has to cross the package, the solder joint, the copper, the laminate and finally the air. Each step is a thermal resistance, and they add in series:
Rtotal = Rpackage + t/(k·A) + 1/(h·A)
The middle term is the one the board controls, and it is worth calculating for 1 cm² of 0.254 mm laminate. With k = 0.20 W/m·K that is 12.7 K/W; with k = 0.69 it is 3.7 K/W; with k = 1.44 it is 1.8 K/W. The last term, natural convection into still air, is about 10 W/m²·K, which for 1 cm² is 1,000 K/W.
That last number is the one that reframes the problem. Convection from a board face is so poor that a board cannot cool itself through its surface at any useful power density: 1 W through 1 cm² of natural convection would need a 1,000 K rise. Heat therefore has to leave through copper — vias, planes, a spreader, a heatsink — and the laminate is an obstacle in the path rather than a route out.
| Path (1 cm²) | Thermal resistance | What that means at 10 W/cm² |
|---|---|---|
| Natural convection from one face | About 1,000 K/W | Not a path at any practical power |
| 0.254 mm PTFE, k 0.20 | 12.7 K/W | 127 K rise: the die is already at its limit |
| 0.254 mm FR-4, k 0.30 | 8.5 K/W | 85 K rise |
| 0.254 mm 4350B, k 0.69 | 3.7 K/W | 37 K rise |
| 0.254 mm 6035HTC, k 1.44 | 1.8 K/W | 18 K rise |
| 0.254 mm copper | 0.007 K/W | Negligible: this is why vias work |
The trade in high-frequency materials is that the additives which raise thermal conductivity also raise Df, and the materials with the lowest loss have the least filler. That is not a rule of thumb; it is visible in the numbers.
| Laminate | Df | k (W/m·K) | R″ at 0.254 mm | Rise at 10 W/cm² |
|---|---|---|---|---|
| RT/duroid 5880 | 0.0009 | 0.20 | 12.7 K·cm²/W | 127 K |
| Taconic TLY-5 | 0.0009 | 0.22 | 11.5 K·cm²/W | 115 K |
| RO3003 | 0.0010 | 0.50 | 5.1 K·cm²/W | 51 K |
| RT/duroid 6035HTC | 0.0013 | 1.44 | 1.8 K·cm²/W | 18 K |
| RO3010 | 0.0022 | 0.95 | 2.7 K·cm²/W | 27 K |
| RO4003C | 0.0027 | 0.71 | 3.6 K·cm²/W | 36 K |
| RO4350B | 0.0037 | 0.69 | 3.7 K·cm²/W | 37 K |
| FR-4 | 0.020 | 0.30 | 8.5 K·cm²/W | 85 K |
Two readings. The first is that 6035HTC is the only common laminate that is both low loss and thermally conductive — Df 0.0013 and k 1.44 — which is why it exists at all, and why it costs what it does. The second is that choosing a material for thermal reasons alone is a mistake in either direction: moving from 4350B to 5880 to gain 0.03 dB/in of loss multiplies the thermal resistance by 3.4, and moving from 5880 to a conductive laminate to gain thermal headroom costs dielectric loss.
The decision rule that follows: pick the loss you need at the frequency and length of the channel, then from the laminates that meet it pick the highest k. Doing it in the other order produces a board that runs cool and does not close its eye, or one that closes its eye and runs at 120 °C.
Since the laminate is an obstacle, the heat has to bypass it, and a via is a cylinder of copper crossing it. The resistance of one via is small but not negligible, which is why they are used in arrays.
| Via (1.6 mm board) | Copper cross-section | R per via | 5×5 array |
|---|---|---|---|
| 0.3 mm drill, 25 µm wall | 0.022 mm² | 185 K/W | 7.4 K/W |
| 0.3 mm, filled with copper | 0.071 mm² | 57 K/W | 2.3 K/W |
| 0.2 mm, filled | 0.031 mm² | 130 K/W | 5.2 K/W |
| Same area of PTFE, no vias | — | — | About 2,900 K/W |
Comparing the second and fourth rows is the whole argument for filled and capped vias: 2.3 K/W against 2,900 K/W is a factor of 1,250, and it is achieved by a process step rather than by a material. The acceptance criteria for that step are mechanical as well as thermal — void content under 2 % and dimple under 25 µm — because a void under a thermal pad is a local hot spot and a dimple becomes a solder void.
Three placement rules matter. Put the array under the thermal pad rather than around it, since lateral spreading in 0.254 mm of laminate is weak. Connect the array to a copper area on the far side, because a via that ends in a signal layer is a stub, not a heatsink. And keep the array pitch tight enough that the copper between vias does not become the bottleneck: at 1 mm pitch the laminate between two 0.3 mm vias is a significant parallel resistance, which is why thermal arrays are often 0.8 mm pitch or closer. The filling and capping process is described in special process holes, and the density side of the same array in HDI PCB performance.
Copper on a laminate does two jobs at once: it spreads heat laterally and it balances the mechanical stress that warp comes from. Both are set by the same number — the residual copper percentage per layer — and both fail in the same way when the number is uneven.
The practical target is 70–80 % residual copper on each layer, with no more than about 15 % difference between layers that are symmetric about the centre. A board that violates that warps during reflow, which lifts one corner of a power device off its thermal pad and turns a 2 K/W interface into a 20 K/W one. The thermal failure is then blamed on the via array, which was fine.
Two design habits help. Use a ground pour on the outer layers around the RF traces rather than leaving them bare, since the pour both spreads heat and shields, but keep it far enough from the trace that it does not change the impedance. And put the copper where the heat is: a symmetric but thin plane under a 10 W device is worse than an asymmetric heavy one that is directly beneath it, because lateral spreading in copper is efficient only within a few thicknesses of the source — the thickness arithmetic is in PCB copper thickness.
Aluminium and copper substrates are the standard recommendation in most explanations, and they are the right answer to a question that most high-frequency boards are not asking. A metal-backed board has a dielectric layer of 60–100 µm between the copper and the metal, and that dielectric is usually an epoxy or a filled polymer with a Df around 0.02 — two orders of magnitude worse than a PTFE laminate. It is a single-sided construction, and it cannot be built as a multilayer RF stack.
| Construction | Thermal path | Df of the dielectric | Where it fits |
|---|---|---|---|
| Aluminium-backed, 0.1 mm dielectric | 0.5–1.0 K·cm²/W | About 0.02 | LED arrays, DC power, single-sided control |
| Copper-backed | Better than aluminium, at 3× the weight and cost | About 0.02 | Very high power density, single-sided |
| PTFE multilayer with via array | 2.3 K/W through the array | 0.0009 | RF front ends, 10 W class |
| 6035HTC multilayer | 1.8 K·cm²/W through the laminate | 0.0013 | Low-loss RF with real thermal load |
The honest summary: a metal-backed board is a thermal solution for low-frequency, single-sided, high-power circuits. It is not a solution for a 10 GHz front end, because the dielectric that makes it thermally good is electrically bad, and because the construction cannot carry the multilayer stack that the RF design needs. Where the RF board does need a metal mass, it is bolted to one — a machined spreader under the board — rather than laminated into it.
The familiar claim that a multilayer board helps thermal management by separating signal layers from heat sources is backwards. Adding a layer adds dielectric between the device and the heat sink, and dielectric is the highest resistance in the chain. If a 10 W device sits on the top layer of a 0.254 mm two-layer board, the heat crosses 12.7 K/W of PTFE to reach the bottom copper. Put four layers under it and the same heat crosses three times that unless the vias go all the way through.
What multilayer does buy is a copper area on which to spread, and a place to put a ground plane that the RF design needs anyway. Both are real benefits, and neither of them is a thermal path by itself. The rule that follows: count the layers for signal integrity and plane continuity, then make the thermal path explicit through vias, rather than assuming that a thicker stack cools better. How the stack itself is chosen is in high-frequency PCB stack-up design.
The same reasoning applies to the lamination step, where the dielectric height that sets impedance also sets thermal resistance — a 0.508 mm core is 25.4 K/W per cm² in PTFE, twice the 0.254 mm value, and it is chosen for impedance rather than for heat. Both effects are worked through in lamination in high-frequency PCBs.
Heat does not only raise the junction temperature; it moves the board. PTFE expands 217–237 ppm/°C in the Z axis, which is 13–14 times copper, and over a reflow excursion of 215 K a 1.6 mm board grows by 75–82 µm. The plated barrel has to stretch with it, and the solder joint at each end has to absorb the difference between the barrel and the component lead.
| Material | Z-axis CTE | Growth over 215 K, 1.6 mm | Consequence |
|---|---|---|---|
| PTFE (5880) | 237 ppm/°C | 82 µm | Barrel strain: needs 20 % elongation and 6:1 aspect ratio |
| PTFE (TLY-5) | 217 ppm/°C | 75 µm | Same, slightly better |
| RO4350B | 32 ppm/°C | 11 µm | Barrel strain is not the limiting factor |
| FR-4, Tg 170 | 50–70 below Tg, 200–300 above | 17–24 µm | Above Tg the rate triples, so the profile matters |
The design rules that come out of it are familiar and are usually applied for the wrong reason: high-elongation plating, an aspect ratio of 6:1 or better, and a thermal via array that is also a mechanical anchor. They are applied because of the expansion, not because of the temperature, and that is why a board can pass a 120 °C bake and fail its third reflow.
Temperature also moves the electrical design, and on a high-frequency board that matters more than the 2 % change in copper resistance. The temperature coefficient of Dk is what sets phase drift, and it has opposite signs in the two material families.
| Material | TCDk | Dk change over 100 K | Phase change at 10 GHz, 100 mm |
|---|---|---|---|
| RT/duroid 5880 | −125 ppm/°C | −1.25 % | About 11° (Dk falls, delay falls) |
| RO4003C | +40 ppm/°C | +0.40 % | About 5.5° in the opposite direction |
| RO4350B | +50 ppm/°C | +0.50 % | About 7° |
Two consequences. A phased array built on 5880 drifts in one direction and on 4350B in the other, so mixing them in one aperture produces a systematic phase error that no calibration table fixes. And a power amplifier that heats its own board shifts the phase of its own output network, which is a slow thermal transient rather than a design error. The phase side of the same physics, including the drift per 100 mm, is in high-frequency PCB signal integrity.
Thermal and electrical requirements pull in opposite directions in five specific places. Knowing the list is most of the design work.
| Objective | Electrical choice | Thermal cost | Resolution |
|---|---|---|---|
| Low loss | 5880, Df 0.0009 | k 0.20, 12.7 K/W per cm² | Accept it, and route the heat through vias and a spreader |
| Thick dielectric for impedance | 0.508 mm core | Doubles R″ | Use a thinner core with a wider line if the loss allows |
| Thin dielectric for density | 0.102 mm core | Lateral spreading collapses | Via array to an internal plane, not surface spreading |
| Metal spreader close to the die | Adds a layer and a via stub | Much better | Back-drill the stub; accept the layer count |
| Copper pour for shielding | Changes impedance if too close | Better spreading | Keep the pour 3h from the trace and stitch it |
1. Heat budget per component, with the output stage listed first. 2. Junction temperature target, and the package Rth that consumes it. 3. Laminate chosen for loss first, then the highest k that meets the loss. 4. Thermal via array under every power pad, filled and capped, 0.8 mm pitch or closer. 5. The array terminated in copper on the far side, not in a signal layer. 6. Residual copper 70–80 % per layer, symmetric within 15 %. 7. Spreader or heatsink interface specified, with flatness and torque. 8. Plating elongation 20 % or better where Z expansion exceeds 50 µm. 9. Phase drift against temperature budgeted where the board carries a phased array. 10. Thermal cycling test defined, because a board that passes at 25 °C tells you nothing about the third reflow.
Ten items, and only items 3 and 4 cost money. The rest are drawing decisions, and they are the ones that get omitted: a via array that lands on a signal layer, a copper balance that warps at reflow, a spreader interface with no flatness callout.
Thermal management on a high-frequency board is not about frequency. The heat comes from the output stage, which on a 10 W RF board contributes about 15 W while the transmission lines contribute under 3 W. The path out is limited by two resistances: 12.7 K/W per cm² through 0.254 mm of PTFE, and about 1,000 K/W through natural convection from the same area. The second number is why a board cannot cool itself through its surface and why every real path is copper — a filled via array at 2.3 K/W per 5×5 group, a spreader, a heatsink.
The material trade is unavoidable but it is narrower than it looks. Of the common laminates, only 6035HTC is both low loss and conductive, at Df 0.0013 and k 1.44; everything else asks the designer to choose which to pay for. What is not a trade is the construction: more layers add dielectric between the device and the sink, and a metal-backed board answers a low-frequency, single-sided question that an RF front end is not asking.
XCE PCB builds 2 to 72 layers on 600×1200 mm panels with filled and capped thermal via arrays, copper balance held within the warp limit, impedance at ±10 % on coupons from the centre and edge of every panel, samples in as fast as 24 hours, inspected to IPC-A-600G and IPC-6015. Send the power map with the stack — the thermal path is decided by the drawing, not by the material name. Request a quote or WhatsApp +86 180 3341 4657. The related property discussion is in performance and role of thermal management in PCB manufacturing.
Send us your Gerber files and requirements for a manufacturability review and a tailored quotation.