MIPI camera works, the board freezes: test the power supply first

A MIPI camera that probes and streams on a second board while this one locks up is not a camera problem. Three OS image versions later, your log shows the same thing on every boot: the HDMI display blanks, drops its cable state, recovers, and repeats — about once a second, for as long as the log runs. That signature is a board browning out under peak load, and the usual culprit is the supply feeding it. This guide walks the diagnosis end to end: read the display-loop pattern in dmesg, measure the power rail under real load, run one power budget that shows why a USB port cannot do this job, and close with a clear call — fix the supply, or return the board.

Read the log: the display retrain loop

Your log carries three pieces of evidence before you touch a screwdriver.

First, the display subsystem is looping. The kernel prints blank - powerdown, then unblank; the HDMI cable state flips Extcon AUX1(HDMI) disable then Extcon AUX1(HDMI) enable; the display re-reads the monitor's capabilities (get RGB quant from EDID) on every pass. Measured from the log's timestamps, blank events land 0.14–0.43 s apart and the full enable, disable, enable sequence repeats every 0.84–0.87 s. The loop runs for the whole captured window — about 8 s in your log — without settling once.

Second, the board stays alive. The dmesg timestamps keep advancing throughout, so the CPU, DRAM and storage never stop. The board is dipping and recovering, not crashing.

Third, the camera is silent. No I2C transaction failure, no sensor probe error, no driver complaint anywhere in the log. A camera fault writes its own lines: the MIPI camera probe failed error -121 guide shows what one looks like. A looping display beside a clean camera log points away from the camera module.

Why a proven-good camera still points at the power rail

Camera bring-up loads a board in stages, and the probe is the cheapest stage. Identifying the sensor over I2C draws milliamps, so it succeeds even on a rail that is already marginal. The heavy load arrives later, when the image pipeline, memory traffic and the display engine all run together. Your log's display line reports a 148.5 MHz pixel clock (nominal-pclk:148500000) — a steady load by itself — and streaming on top of it pushes current demand to its peak. A supply near its limit lets the rail dip, the HDMI transmitter drops the link, and the retrain loop begins.

Put numbers on the source side. A USB 3.0 data port delivers at most 0.9 A at its nominal 5 V — 4.5 W, the ceiling the USB 3.0 specification sets for a port. Cable length eats into that before it reaches the board: the power conductors of a thin cable run about 84 mΩ/m (24 AWG), so 1 m is roughly 0.17 Ω round trip, dropping 0.34 V at 2 A. A port at that nominal voltage behind 1 m of thin cable lands near 4.66 V at the board — below the 4.75 V floor before the supply itself sags at all. The camera power decision between a USB port and an external DC supply covers the same budget for steady streaming; a freezing board is that budget failing hard.

Likely causes in order, and the test that clears each

Given a camera that streams on other hardware and three clean OS installs, rank the suspects like this:

1. The supply cannot deliver peak current. Run the board from a bench supply with a current readout, or from a DC barrel adapter rated 4 A at the board's input voltage. Watch the readout as the camera starts. A peak above the old source's rating, plus a loop that stops, closes the case.
2. The input path wastes voltage. A long or thin cable, an unpowered hub, a worn jack. Connect the DC supply straight to the board's power jack with a short lead and measure at the jack. A 20 AWG lead (about 33 mΩ/m) drops 0.13 V at 2 A over a metre; 24 AWG drops 0.34 V. Short and thick wins.
3. The display link loops on its own. Swap the HDMI cable and the monitor. Then run headless from the serial console: if blank and unblank lines keep scrolling with no monitor attached, the display hardware is cleared.
4. Software. Three clean OS images plus a camera that probes on identical hardware make a driver fault unlikely. If your boot log shows probe errors instead of a clean bind, that is a different failure — the camera I2C communication failure guide covers those signatures. Reflash only when the log asks for it.
5. The board itself. The cause you reach only after 1 through 4 come back clean. The next two sections turn this into a measured decision and the warranty call.

Worked example: a USB 3.0 port against a 4 A barrel supply

Inputs, each with its source:

• Source A — a USB 3.0 data port: 0.9 A, the specification ceiling for the port (4.5 W at the nominal rail voltage).
• Source B — a 4 A barrel supply, a standard rating for this board class (20 W at the same voltage).
• Floor — 4.75 V at the input jack under peak load, the 5% low tolerance edge of the nominal rail.
• Symptom to explain — the display retrain loop repeating about every 0.85 s, from your log.

The rule in words: a source holds the rail when its rated power covers the board's peak demand, and power is voltage times current.

Arithmetic:

• Port ceiling: 5 V × 0.9 A = 4.5 W.
• Barrel supply: 4 A at the same rail = 20 W.
• Headroom: 20 W ÷ 4.5 W ≈ 4.4.

What moves the numbers: a USB-C port advertising 1.5 A lifts the ceiling to 7.5 W, and a 3 A port to 15 W — better, still under 20 W. Cable loss subtracts from whatever the source makes: 2 A through 0.17 Ω of round-trip cable dissipates (2 A)² × 0.17 Ω ≈ 0.7 W and drops about 0.34 V at the board.

Result: the port tops out at 4.5 W — 15 W at best, negotiated — while the barrel supply holds 20 W.

Decision: if the loop and the freeze reproduce on port power and disappear on the barrel supply with the jack voltage holding 4.75 V or above under load, the supply was the fault — keep it and close the case. If the loop persists with 4.75 V held at the jack, the supply is cleared, and with the camera proven elsewhere the board itself is the fault.

The fix: size for peak, measure under load

Size against peak, with headroom. Take the board's documented peak current at its DC input, add the camera module and every USB peripheral, then buy at least 1.5× that total. For the board class in your log that lands at a 4 A barrel supply — 20 W — rather than any port or charger. Use a regulated switching adapter with the right plug size and polarity, straight into the board's jack on a short lead.

Then measure once: an inline USB power meter or the bench supply's readout during camera start gives you the true peak, and a meter at the jack confirms the 4.75 V floor while the display and camera run.

The warranty decision: prove the supply, then the board

Send nothing back until the supply swap has run. A board returned on a marginal power set-up comes back with no fault found and costs weeks.

Your evidence pack has four lines:

• The camera module and its ribbon stream on a second board — one swap clears the sensor, the module and the cable, the same method as the swap test that isolates a faulty FFC cable.
• Three clean OS image versions installed, driver binding clean in each boot log.
• The freeze and the display loop reproduce on the original power source.
• The behaviour persists with a supply rated above the board's documented peak holding 4.75 V at the jack under load.

If the loop disappears on the proper supply, the supply was the fault and the board is fine. If it persists, the board is the fault: request the return and send this evidence with it. Field history backs the order of suspicion — a parcel-scanning camera returned to service by replacing the camera's power supply alone, with no camera fault involved. If you want a second opinion on a supply rating before you order, contact us.

When the board lives in a machine

A bench build can hang off a barrel adapter; a machine cannot. Cabinets distribute 24 V or 48 V and derive the board's low-voltage rail locally through a step-down converter sized to the same budget, peak plus 30–50% headroom, with the input fused. An enclosed industrial supply makes a durable front end for that chain. The LRS-75-48, for example, is a 48 V, 76.8 W enclosed unit built for exactly this duty.

Fuse the converter input, keep the 5 V lead short, and confirm the jack voltage once on the first power-up. The power supplies category holds the enclosed units, and the power and accessories finder filters them by type.

Prevention: budget the rail at design time

Budget at design, not at failure. List the board's documented peak current, the camera module, the display and every USB device; add 30–50% for inrush and supply aging; buy to that number. On the first bench run, log the minimum jack voltage during camera start and accept nothing below 4.75 V.

Boards that derive sensor rails from the main input add a second exercise: keeping the 3.3 V and 1.8 V loads balanced so one hungry peripheral cannot drag its neighbours — balancing 3.3 V rails in embedded systems covers that in depth. When you pick the module itself, match its current draw to what the board can feed: the board-level camera selection guide walks that choice, and the GMSL camera power budget guide shows the same input-and-idle-current arithmetic for a capture card.

Before you order: five checks

• Voltage and polarity — match the board's DC input rating and its marked centre polarity exactly; a board built for a 5 V input does not forgive 9 V.
• Current rating — at least 1.5× the measured peak, not the idle figure the data sheet prints first.
• Regulation — a regulated switching supply only; an unregulated adapter sags with load by design and recreates the fault you are fixing.
• Connector — the plug's outer diameter and inner pin must fit the board's jack; order the supply with the correct plug or a known-good lead.
• Proof plan — include an inline USB power meter so the first run measures jack voltage under load, not a guess.

Run the five checks, fit the supply, and watch one camera start. If the jack holds 4.75 V and the loop is gone, you fixed it for the cost of an adapter. If the loop survives a supply that holds its floor, the evidence pack above is complete — send the board back. The cameras topic page collects the rest of the camera-side diagnostics, from probe errors to link drops.