serial-bench one bench, five serial buses - read-only archive

serial-bench / bench

The bench

Updated 2026-08-21 · one desk, one shelf, no climate control

Every number on this site came off one of these. I list the specifications because a measurement without its instrument's limits is an opinion, and because three of the six write-ups here turned on a limit rather than on a fault.

I deliberately do not name manufacturers or model numbers. What decides whether a measurement is admissible is the sample rate, the analogue bandwidth and the threshold behaviour, not the badge on the front panel.

Instruments, in the order I reach for them

Instrument	Key figures	Good for	Not for
Logic analyser	8 channels, 24 MS/s, 41.7 ns interval, 1.5 V fixed threshold	protocol order, bit values, addresses, frame structure up to ≈3 MHz	edge positions, duty cycle, anything analogue
Oscilloscope	2 channels, 100 MHz bandwidth, 1 GS/s, 8-bit vertical	rise times, overshoot, levels, reflections, 1 ns-class edges	wide buses, long protocol sequences
Differential probe	70 MHz, ×10 and ×100, ±140 V common mode	CAN pairs, RS-485 pairs, ground-to-ground potential	single-ended millivolt work
Bench supply	0–30 V, 0–3 A, current limit settable to 10 mA	finding shorts before they find the board, sink-current tests	transient load response
Isolated DC-DC bricks	5 V in, 5 V out, 1 W, 2.5 kV isolation	separating a remote transceiver from its local ground	anything above a few tens of milliamps
Resistor and capacitor decade boxes	1 Ω–100 kΩ in 1 Ω steps; 10 pF–1 µF	sweeping pull-up values without a soldering iron	high-frequency substitution, the leads dominate

Probing habits that changed my results

1. Ground spring instead of the ground lead. The crocodile lead on a scope probe adds enough inductance to invent overshoot. On the CAN work it fabricated a 600 mV ringing artefact that disappeared the moment I used a spring tip. If a waveform shows ringing, re-probe before you believe it.
2. Trigger on the thing you doubt, not on the clock. Framing errors are rare events; triggering on chip select or on a start edge and letting it sit for a few minutes catches more than a thousand screenshots of healthy traffic.
3. Record the capture settings in the same file as the capture. Sample rate, threshold, probe attenuation. I lost an entire session's worth of credibility once because I could not say afterwards whether a run had been at 24 MS/s or at 8.
4. Measure the same quantity a second way. The SPI mismatch was proved by reading the same register over the sensor's other interface. One instrument agreeing with itself is not corroboration.

What I do not have, and what that costs

No frequency counter or calibrated reference

This is why the baud-rate measurement ends with an unresolved 0.5 % discrepancy. I can measure a ratio between two signals on the bench; I cannot tell you the absolute frequency of either to better than my analyser's quantisation.

No time-domain reflectometer

Cable impedance and velocity factor are taken from the cable's published figures rather than measured. On the 95 m CAN segment I used 5 ns/m as given, and said so in the post.

No climate chamber

Everything here is at room temperature, roughly 21 °C. Crystal drift over temperature, which matters a great deal for baud-rate budgets in the field, is therefore outside what I can demonstrate.

Only two analogue channels

I cannot show you a differential pair and its chip select simultaneously with analogue fidelity. Where a post needed both, the two captures were taken separately on the same trigger condition, which is weaker evidence and is flagged where it happens.

← #1 - why the sampling ratio decides everything all messages →