When you send a high-speed signal down a PCB trace or cable, the wire acts like a low-pass filter. High frequencies attenuate more than low frequencies — a 5 GHz harmonic on a 10 Gbps lane might lose 10 dB while DC loses nothing. By the time the signal reaches the receiver, sharp edges have smeared into rounded blobs and adjacent bits bleed into each other (inter-symbol interference, or ISI). Equalization fixes this at the receiver. Pre-emphasis fixes it at the transmitter — and it's often cheaper.

The idea: deliberately distort the signal at the TX in a way that exactly cancels the channel's distortion. Since the channel attenuates high frequencies, the TX boosts high frequencies before sending. Equivalently — and this is how it's usually described — the TX de-emphasizes the low-frequency content. A bit that holds the same value as its neighbors (a "run") gets driven at a lower voltage swing than a bit that just transitioned. Transitions get the full swing; steady-state bits get cut.

The implementation is a 2-tap FIR filter built from current-mode drivers:

• Main tap drives the current bit at full strength (say, +1.0)
• Post-cursor tap drives the previous bit's value inverted at reduced strength (say, -0.25)
• Output = main + post-cursor, summed in current at the differential pads

When two consecutive bits are the same (1→1), the post-cursor subtracts, so the second bit drives at 0.75 instead of 1.0 — de-emphasized. When bits differ (0→1), the post-cursor adds in the same direction, boosting the transition to 1.25 — pre-emphasized.

Concrete example: PCIe Gen3 (8 GT/s) over 20 inches of FR4 sees about 15 dB of loss at the Nyquist frequency (4 GHz). The PCIe spec defines 11 preset coefficient values the TX advertises; the link partner picks one during equalization training. Preset P4 uses roughly -3.5 dB of de-emphasis with no pre-shoot — typical for moderate channels. Without it, the eye closes completely and the BER explodes from 1e-12 to essentially 1.

Rule of thumb: de-emphasis in dB should roughly equal the channel's loss at Nyquist. For 6 dB of channel loss, use 6 dB of de-emphasis (post-cursor coefficient ≈ 0.5 of main). Overshoot the equalization and you over-boost noise; undershoot and ISI wins. Modern SerDes link-train these coefficients automatically, sweeping values and watching the BER.

Pre-emphasis is cheaper than RX equalization because it operates on clean, known data — no noise to amplify. The TX knows exactly what bit it sent last. The RX, by contrast, has to guess from a noisy waveform.