DRAM Detective · Case File 2.0: The Sound Engineer

Published 2026-11-19 by Peter Hu

PHY is the analog front-end inside the controller — the translator between digital and analog.

It can fix reversible degradation, but not irreversible damage: while the eye is open, it can amplify, center, and straighten; once the eye closes, the information is gone and no knob helps. So PHY is the after-the-fact sound engineer; SI is the up-front studio builder.

Auto-plays · open full screen ↗

One — what PHY does

PHY (Physical Layer), in a memory interface, is the analog front-end on the controller chip.

The controller only speaks 0 and 1, but what runs down the trace is a voltage waveform. On transmit, PHY turns 0/1 into a waveform; on receive, it judges the waveform back into 0/1. The two sides don’t share a language — PHY translates.

Two — five jobs: transmit, receive, timing, compensate, calibrate

Job What it does
transmit (TX) 0/1 into an electrical signal, driving DQ/DQS/CK, controlling drive strength and slew
receive (RX) the signal judged back into 0/1 — Vref decides, DQS Gate sets the sample instant, CTLE/DFE equalize
timing PI/DLL make the precise phase and clock that decide when to sample
compensate ODT absorbs reflections, CTLE/DFE restore loss, DCC fixes duty cycle, de-skew removes skew
calibrate ZQ impedance calibration, rank delay storage

These five jobs spread out into the eight white-box chapters that follow: 2.1 PI, 2.2 DLL/PLL, 2.3 DQS Gate, 2.4 Vref, 2.5 the equalizer trio, 2.6 transmit driver, 2.7 DCC, 2.8 rank timing storage. The next chapters open each knob and see how it’s built.

Three — the boundary: is the eye still open?

PHY can fix reversible degradation, but not irreversible damage.

SI problem Can PHY fix it? Why
mild tail / reflection / crosstalk / jitter yes the signal is still there, just blurred — equalization, centering, de-skew can widen the eye
eye fully closed (loss too high, signal drowned in noise, eye height zero) no the 0/1 information is already lost; amplifying only amplifies the noise too
resonant notch (a channel frequency carved out) no structural damage in the frequency domain, unrecoverable in time
power noise / ground bounce (bad PDN) no a power problem, not a trace problem — equalization can’t save it
bad contact / oxidized gold fingers no no signal arrives at all

One criterion: while the eye is open — even 1 mV — PHY can amplify, center, and straighten it; once eye height and eye width are both zero, the information is gone, and any amplification just lifts the noise along with it.

Four — at the limit, go back and fix the board

PHY tunes after the fact; SI builds the studio up front. A bad studio — reverb, crosstalk, power noise — no engineer can save an already-ruined recording. At that point, stop turning PHY knobs and fix the source:

  1. Lower the frequency — the most direct fallback. Drop to a frequency where the eye reopens; trade speed for margin.
  2. Rework the traces: shorter runs, wider spacing, matched impedance, ground isolation, better layer changes.
  3. Reduce load: fewer ranks / DIMMs, and the reflections and loss come down with them.
  4. Fix the power (PDN): more decoupling caps, better power planes, to cure ground bounce.
  5. Swap the die / DIMM: stronger drive, lower load.
  6. Fix the contact: clean the gold fingers, reseat the slot — the “can’t find DIMM” class of failure.

None of these six is something training can do for you — they all go back to the board and the parts.

While the eye is open, PHY is the sound engineer; once the eye closes, go back and build the studio.