DRAM Detective · Case File 0.8: The Label, the Ruler, and the Subtraction

Published 2026-11-19 by Peter Hu

Timing numbers don’t all come from one place: some are read from the SPD and fixed, some are measured by training after power-up, and some are computed from a spec baseline plus training results.

Know which hand manages each number, and you’ll know which knob training is turning — and why.

Auto-plays · open full screen ↗

One — three kinds of numbers, three origins

Origin When it’s settled Typical Written by
configured (static) written into the SPD at the factory / looked up from spec, read at power-up tCL / tRCD / tRP / tWR / tRFC / tREFI / tCCD_L the DIMM vendor
measured (dynamic) measured by training after power-up DQS/DQ delays, sample phase, gate timing, Vref, RTT, DFE taps the MRC training
computed (mixed) after training, from a baseline plus results turnaround, rank-to-rank delta the MRC’s arithmetic

The label, the ruler, the subtraction. Configured numbers are the spec stamped on the nameplate — copy them. Measured numbers come from a tape measure on-site. Computed numbers are two measurements subtracted — no single ruler can read that gap directly.

Two — the full map

Configured: fixed at the factory, never trained

These record the die’s own physics. The vendor measures them, writes them into the SPD (or they’re looked up from the spec), and they’re read into the controller at power-up and never touched. The reason they aren’t trained is simple: they don’t depend on how the board is routed, only on the die’s speed and internal circuitry. Later chapters barely mention them again — training turns the “measured” category — so here’s the full story in one place.

The four waits in the read/write flow (the four steps from Case 0.0):

The two refresh ones:

The rest, all about commands fired back to back:

And one that’s neither in the SPD nor given by training, only looked up from a table:

Measured: measured after power-up

This is where training actually works — numbers nobody could know before the board was assembled.

Knob What it trains
CA de-skew command/address alignment
gate timing (RcvEn) when the DQS gate opens
tDQSS write DQS aligned to CK
DQS delay + sample phase read DQS arrival + sample point
per-bit DQ delay per-bit de-skew
DFE taps the ISI-cancelling tap coefficients
RTT / ODT setting termination impedance
CTLE / DXC treble boost + crosstalk cancellation
duty-cycle DCC fixing DQS/CLK duty cycle
Ron / Comp transmit-side drive impedance
slew + Tx EQ edge slope + pre-emphasis
Vref the decision line
rank-to-rank switch offset the cross-rank switching offset

Computed: baseline and training results put together

This kind has no single true value to measure — only arithmetic.

Quantity How it’s computed
Turnaround (Wr2Wr / Rd2Rd switch gap) JEDEC tCCD baseline + first-to-last rank timing delta + guardband
delta (first-to-last rank timing gap) max minus min of the timing settings across all ranks

Three — why three kinds at all

The configured kind needs no training. tRCD, tCL, tRP are set by the die’s speed and internal circuitry; the board’s routing, temperature, and voltage can’t change them. The vendor measured them and wrote them into the SPD — reading them at power-up is enough, no point measuring again.

The measured kind can only be measured on-site. Trace length, temperature, voltage, and die batch all vary, and — worse — the combination of “this die on this board” is something nobody knows at the factory. So after power-up, training drives traffic, watches the feedback, and measures each alignment value one by one.

The computed kind has no choice. It wants a relationship between two quantities — turnaround is the difference between the first and last rank’s timing, and there’s no register holding a “true value” waiting to be read. You subtract the two trained numbers, then stack the spec baseline tCCD and a guardband on top.

So there’s a crude but useful trick for reading a training log: first recognize which knob in the “measured” category this step is turning, and you know which line it’s serving. Training spends all its effort on that category — the configured and computed kinds are the two ends, nailed down.

Copy the label, measure with the ruler, subtract the gap — training always turns the ruler.