Dielectric constant and loss tangent

In short: A laminate's dielectric constant (Dk) sets how fast your signal travels and the impedance of every trace, while its loss tangent (Df) sets how much of that signal turns into heat — so above a few GHz you trade cheap FR-4 for low-loss materials like Rogers because Df, not copper, dominates the loss budget.

Every high-speed interconnect you route — a DDR5 byte lane, a 25 Gbps SerDes channel, a PCIe Gen6 link — lives inside a dielectric. Two numbers describe that dielectric and quietly decide whether your link closes: the relative dielectric constant (Dk, or ε_r) and the dissipation factor (Df, or loss tangent, tan δ). Get them wrong and your impedance is off, your eye is closed, and your board fails at the lab bench no amount of equalization can rescue.

What Dk actually does to your signal

Dk is the ratio of a material's permittivity to that of vacuum. It controls two things an SI engineer cares about: propagation velocity and characteristic impedance. Signal velocity is v = c / √Dk_eff, so on FR-4 with an effective Dk around 3.8–4.2, a signal crawls at roughly 15 cm/ns versus 30 cm/ns in free space. That directly feeds your length-matching: a 1 mm skew between DDR traces is about 6.5 ps on FR-4, and DDR5 write-leveling windows are measured in single-digit picoseconds.

Dk also sets impedance. For a microstrip, raising Dk lowers Z₀ for a fixed geometry, which is why your stackup engineer specifies Dk to two decimal places. A controlled-impedance 50 Ω single-ended trace designed for Dk 4.3 but fabricated on a lot that measures 3.9 will come out several ohms high — enough to blow a ±10% spec and bounce the board. Field solvers in Cadence Sigrity, Siemens HyperLynx, or Polar Si9000 all take Dk as a primary input, so the number you feed them has to be the real, frequency-correct value, not the 1 MHz datasheet headline.

Loss tangent: where the energy goes

Df quantifies dielectric loss — the fraction of stored field energy dissipated as heat each cycle. Dielectric attenuation scales roughly as α_d ∝ f · √Dk · Df, so it climbs linearly with frequency. That linear term is the killer: at 1 GHz copper (conductor) loss dominates, but by 10–28 GHz dielectric loss on a lossy laminate overtakes it and sets the slope of your insertion-loss curve.

Standard FR-4 has a Df around 0.02; a premium low-loss laminate like Rogers RO4350B sits near 0.0037, and ultra-low-loss PTFE systems (RO3003, Megtron 7) reach 0.001–0.002. At 28 GHz that is the difference between a channel that closes and one that does not.

FR-4 vs Rogers: a working comparison

Notice two things beyond the headline Df. First, Rogers laminates hold a tighter Dk tolerance (often ±0.05) versus FR-4's loose ±0.2, which matters enormously for controlled impedance and for phase-matched RF feeds. Second, FR-4's Dk and Df drift with frequency, temperature and resin content, so a trace that is 50 Ω at 100 MHz may not be at 10 GHz. For a deeper look at how geometry interacts with the dielectric, see our guide on microstrip vs. stripline — stripline buries the trace fully in dielectric, so its effective Dk and loss behave differently from microstrip.

Rule of thumb: if your Nyquist frequency (≈ half the bit rate) exceeds ~8–10 GHz, or your total channel is long, FR-4's dielectric loss will eat your margin. That is the usual trigger to move to a low-Df laminate — or to shorten and re-plan the channel.

Glass weave, roughness and the "effective Dk" trap

Real laminates are woven glass in resin, and glass (Dk ≈ 6) and resin (Dk ≈ 3) are not uniform. A trace running over the glass bundle sees a different local Dk than one over the open resin window — this is the infamous fiber-weave effect, which creates intra-pair skew and mode conversion in differential pairs. Mitigations include spread/mechanically-flattened glass (1078/3313 style), routing at a 7–10° zigzag, or specifying low-Dk glass. Meanwhile copper roughness, added to improve adhesion, lengthens the current path and inflates apparent conductor loss; modern solvers use Huray or Hammerstad models to capture it. The practical lesson: the "effective" electrical Dk your channel sees is a system property, not a single datasheet cell.

Designing with the numbers

A disciplined flow looks like this. Pull the laminate's Dk/Df-vs-frequency table (Bereskin or clamped-stripline data), not the 1 MHz value. Feed the frequency-correct Dk into your 2D field solver to hit impedance, and the Df into a 3D/channel simulator — Cadence Clarity or Sigrity, ANSYS, or Siemens HyperLynx — to build the insertion-loss budget against your SerDes IBIS-AMI model. Budget conductor loss, dielectric loss, via stubs and connector loss separately; dielectric loss is the one that scales with length and frequency, so it drives your material decision more than any other. If simulation shows you are 3 dB short at Nyquist on FR-4, the fastest fix is usually the laminate, not another CTLE tap.

Teams building boards that must pass compliance often validate stackups against vendor Dk/Df curves and TDR measurements before committing to a fab. If you need that work done rather than learned, Avecas offers high-speed PCB and signal-integrity design services. To go deeper on the simulation side, our companion article on building an insertion-loss budget walks through the full channel math.

Want to master stackup design, Dk/Df selection and channel simulation end-to-end? CourseTron's high-speed PCB and signal-integrity track takes you from microstrip math to closing a 25G eye in HyperLynx. Explore the full course catalog and start building boards that pass first time.

Frequently asked questions

Can I just use FR-4 for everything and add more equalization?

Up to a point. Equalization recovers lost amplitude but amplifies noise and crosstalk along with signal, and it cannot fix the steep slope that high Df imposes on long channels. Past roughly 10 GHz Nyquist, a lower-Df laminate buys cleaner margin than any amount of EQ.

Why does the datasheet Dk not match what my board measures?

Datasheets often quote Dk at 1 MHz or a single frequency and under ideal conditions. Real Dk drops with frequency, varies with resin content and glass style, and the effective value depends on your trace geometry. Always use frequency-dependent data and, ideally, a measured TDR/VNA correlation.

Does a lower Dk always mean lower loss?

No. Dk sets velocity and impedance; Df sets loss. They are independent properties. A material can have moderate Dk but very low Df (good for loss) or vice versa. For loss, look at Df; for timing and impedance, look at Dk.

What is the difference between Df and tan δ?

They are the same quantity. "Dissipation factor," "loss tangent," and "tan δ" all name the ratio of the lossy (imaginary) to the storage (real) part of the complex permittivity.