How to fix “power supply voltage is unstable” on Bitmain Antminer S17 ASICs

On an S17, power supply voltage is unstable is not a complaint about your wall outlet — it is the control board reporting that the voltage it commanded the APW9 to produce and the voltage it measured coming back no longer agree. The S17 runs BM1397 chips, 48 per hashboard, wired as 12 voltage domains of 4 chips at roughly 1.55 V per domain — an ~18.5 V series stack that the board’s own boost converter (21–24.5 V output) feeds from the PSU’s adjustable 14.5–21 V rail. When that rail sags or refuses to track the setpoint, firmware drops the boards rather than cook 144 chips. What it costs you: a machine that hashes for minutes and drops, and a repair bill that doubles if you keep power-cycling a board that is shorting the rail.

What the error actually reports

The S17 and S17 Pro ship with the APW9; the S17+, S17e and T17+ use the APW9+. Both are adjustable: 14.5–21 V DC on the main rail (170 A on the APW9, 200 A on the APW9+, 3600 W at 220 V), plus a separate fixed 12.3 V / 12 A rail. That second rail is why you can still reach the web UI while the hashboards are dark.

The control board talks to the PSU over a 4-pin signal cable carrying SDA, SCL, EN (active LOW) and GND. It writes a setpoint, then reads back — and its two reads are not the same thing:

• GET_VOLTAGE returns the stored DAC setpoint — a number firmware itself wrote. It is not a measurement.
• MEASURE_VOLTAGE returns an ADC reading of the actual output. Allow 1–3 s of settling after a setpoint change before it means anything.

“Unstable” is the measured value failing to track the setpoint. Neighbouring log lines say which failure you have: ERROR_POWER_LOST and get power type version failed! mean the PSU never passed model-string validation (firmware whitelists the PSU’s 16-byte ASCII model string), whereas a rail that starts correct and collapses under load is a current problem, not an identity problem. Pull the log first — see reading the S17 kernel log.

Likely causes, ranked

1. The AC feed cannot hold the load. The APW9 has two AC inlets and is rated 3600 W on 200–240 V. It will run on 110/120 V, but at heavily reduced output — a 120 V feed, one populated inlet, or a shared 15 A branch circuit will pull the rail down every time the tuner ramps.
2. A hashboard is dragging the rail down. A shorted domain or chip is a low-impedance load: the PSU current-limits, measured voltage falls away from setpoint, firmware calls it instability. The fault is on the board, not in the supply.
3. Degraded PSU. Aged bulk capacitors push ripple high enough that the measured rail can no longer track the setpoint.
4. Contact resistance. At 150 A+ a discoloured 6-pin, a spread female pin or an oxidised contact drops real volts under load and looks perfect at idle.
5. The 4-pin signal cable. Loose, corroded or crushed — the control board loses readback and cannot verify the rail.
6. Watchdog trip. The PSU’s I2C watchdog cuts the main output after roughly 60 s without a heartbeat (firmware feeds it every 10–30 s); the 12.3 V rail stays up, so the miner is alive to complain. A hung mining process looks exactly like a power fault.

Diagnose before you desolder anything

• Move it to a known-good 240 V circuit, both AC inlets fed.
• Swap in a known-good APW9/APW9+. If the error follows the PSU you are done; if it stays with the chassis, the fault is downstream. Removal procedure: detaching the APW9 from an S17/S17 Pro.
• Isolate by board. Reconnect boards one at a time and watch which one collapses the rail. Some firmware builds refuse to start with fewer than three boards attached — that refusal is not a new fault.
• Reseat the 6-pin power and 4-pin signal cable and inspect both for browning. Reconnect in Bitmain’s order: negative first → positive → signal cable; reverse to disconnect.

Electrical measurements

Safety: the low-voltage DC domains on an S17 hashboard are not a shock hazard. The PSU’s mains side is — its PFC capacitors can hold roughly 400 V after you unplug it. Never probe inside an energised APW9.

1. AC at the inlet, under load. Measure while the miner ramps, not at idle. More than a few volts of sag is a supply-side problem you cannot fix on the bench.
2. Main DC rail at the 6-pin, under load. Black probe to a ground pin, red to a positive pin on the same connector. Expect the commanded rail — mid-to-high teens climbing toward the low twenties — not 12 V.
3. Drop across the 6-pin. Probe PSU-side and board-side of the same pair while hashing: tens of millivolts is normal, hundreds means a burned contact.
4. Ripple, on a scope. AC-couple across the rail. Bitmain’s APW-series ripple spec is <120 mV / <1%; treat that as a guideline on the APW9. Hundreds of millivolts is a PSU you retire, whatever the DMM says.
5. Domain sweep, per board. Power the board on a fixture, wait 2–5 minutes to stabilise, meter on millivolt resolution, black probe to board ground (not the heatsink — short risk), red probe on each of the 12 domain test points. Healthy: all domains within ±50 mV of each other, around 1.55 V each. 100 mV+ below average = partial short (failed chip or blown cap). Well above average = open circuit.
6. Boost output. The S17 boosts its rail into the 21–24.5 V range. No boost, no domain voltages, zero chips found.

Board-level repair

• Near-zero domain. Find the short: a failed BM1397, a shorted LDO, or a punctured decoupling capacitor inside that domain. Current-limit the rail from a bench supply and find the hot component with an IR camera.
• Open domain. Domain voltage climbs because nothing draws current. Look for a cracked joint or lifted pad, typical of a hard-cycled board.
• Ambiguous chip fault. Use the dichotomy method: short RO to 1.8 V at the midpoint chip, see which half still answers, bisect again.
• Burned 6-pin socket. Replace the connector; do not clean it up and reassemble. Sockets and LDOs are stocked as ASIC repair parts.

Verify the fix

• All three boards enumerate a full 48 chips each — 144 total. A partial count means you fixed the rail and left a chip fault.
• Measured output tracks the setpoint through a full tuning ramp. Tuning values are calculated at runtime, so the rail works hardest right after a restart.
• Four hours minimum at the machine’s real duty cycle. Instability that only appears once the room warms will not show in a ten-minute bench test.

Common mistakes

• Chasing 12 V on the hashboard connector. The 12.3 V rail feeds the control board and fans; the hashboards run off the adjustable 14.5–21 V rail.
• Defeating the watchdog to make the error go away. It exists so a hung host cannot leave 150 A sitting on a shorted board.
• Fitting a non-APW9 supply. No valid model string on the I2C bus means ERROR_POWER_LOST — an APW3 or APW7 has no I2C MCU to answer at all, however good the supply is.

Related

• “Hashboard voltage is inconsistent” on the S17+ — the domain-level sibling fault.
• APW9/APW12 PSU problems and solutions
• ASIC Fault Finder
• Start a repair