Cold does not kill a miner by being cold. It kills it three other ways: condensation when a sub-zero board meets humid indoor air, thermal cycling that cracks BGA solder joints and the domain-boundary resistors, and cold-start behaviour in the fans, PSU and firmware that operators mistake for a fault and then “fix” with repeated restarts. A board that reaches our bench with a chain break after a Quebec winter is usually not a chip failure — it is an operator who power-cycled a −20 °C machine six times.
What cold actually does to the hardware
- Condensation. Water forms on the coldest surface in the room. A miner cold-soaked at −20 °C and carried into a 20 °C shop at 50% RH is far below the dew point, so moisture plates out on the PCB, under the heatsinks, and on the temperature-sensor ICs on the bottom side. Powering up wet buys corrosion — an open domain or a chain break months later.
- Thermal cycling. Every cold start drives a large ΔT through the board in minutes. It is the named root cause of two failure modes we see constantly: BGA solder crack (intermittent chip dropout) and cold solder joints on the CLK coupling resistors at domain boundaries, where the chain breaks exactly at a domain edge. A cold joint on a CI/BO resistor is a different signature — it breaks the chain mid-domain.
- Firmware thermal logic. The protection loop is built for heat, not cold: above roughly 75 °C it holds the chips in reset, drops the domain rails to 0 V, waits for the board to fall below 45 °C, then restarts about 100 MHz lower. Bouncing through that loop is self-inflicted thermal cycling, and a stalled cold fan trips it as fast as a hot aisle.
Receiving a miner in winter
- Do not open the box. The sealed carton and bag are your vapour barrier — let moisture condense on the outside. Open it only once the chassis is no longer cold to the touch.
- Budget hours, not minutes. A 19- or 21-series chassis is kilograms of aluminium heatsink. Off a truck at −20 °C it needs several hours in a dry room; low RH matters more than high temperature.
- Shake test before installing. With the miner off, tilt and gently shake it. A rattle means a heatsink detached in transit — a bare die goes straight into thermal shutdown. Do not power it; open it and reseat with fresh paste (Fujipoly SPG-30B).
Cold-starting a machine that has been sitting
- Recirculate exhaust into the intake — angle the miner, or duct hot-aisle air into the cold aisle — until intake reads above 0 °C.
- Warm the intake face evenly. A heat gun into one corner produces the localised ΔT that cracks solder.
- Start at reduced power. Autotuners calculate frequency and board voltage targets at runtime (they are not presets), so a cold first start ramps on its own — give it that time.
- Do not click Restart on a machine flagged abnormal during a cold snap. Fix intake temperature first: repeated restarts are the damage mechanism, not the diagnostic.
Fans, PSU and the aux rail in the cold
- Fan bearing grease stiffens below freezing. A fan that cannot reach commanded RPM trips a fan fault and the firmware refuses to hash — then clears once the bearing warms. Check before condemning it: Antminer fan troubleshooting.
- On an APW12, fans and control board share the second output (12 V, 15 A); hashboards run off the main 12–15 V / 233 A rail — a jammed fan loads the rail the controller lives on.
- Input voltage matters in cabins and garages. APW12 and APW17 are 200–240 V input only — they will not run on a 120 V circuit, and no firmware flag changes that. APW3/APW3++, APW5, APW7, APW8, APW9 and APW9+ do run on 110/120 V at reduced output: an APW3++ on 120 V/15 A delivers roughly 1,200 W derated.
Electrical checks after a cold event
Safety: the domain rails sit between ~0.32 V (S19 Pro) and ~1.2 V (S21) — not a shock hazard. The PSU is: its PFC bulk capacitors sit at 410–420 V DC and stay charged after you unplug it.
- Inspect first, power second. Under a 10× loupe look for water marks, green/white corrosion, and cracked solder on the domain-boundary resistors.
- Power on and wait 2–5 minutes for stabilisation before trusting any reading.
- Measure the domain test points. Healthy boards hold every domain within ±50 mV of the others; any domain off by more than 100 mV is a fault. Voltage here is regulated per domain, never per chip.
- Check boost output: 19–25 V by model (S19: 14 V → ~19 V; S19 Pro: 12.6 V → ~20 V via Q9; S21: → ~25 V via U206). No boost means no domain voltages and 0 chips found.
- Read chip temperatures. All chips should sit within ±10 °C of the board average. One cold chip among hot neighbours is a dead chip; a whole cold region is a domain not powering up.
- If temperature reads are missing or absurd, scan the I2C bus — sensors sit at 0x48–0x4F (LM75A, TMP1075, NCT218, T451), under the heatsink on the bottom side. A stuck SDA/SCL line trips thermal protection falsely: temperature sensor fault.
Reading a winter chain break
The chips are daisy-chained: CI/CO, CLK, RST and BI/BO hand off from one chip to the next. When a chip dies — or a whole domain fails to power up — everything downstream of the break goes invisible to the control board. The reported ASIC count therefore does not count the dead chips; it locates the break, which sits immediately after the last detected chip.
- S19 (BM1398): 76 chips per board, 38 domains, 2 chips/domain. A count of 74 means the chain stopped at the very last domain.
- S19 Pro (BM1398): 114 chips per board, 38 domains, 3 chips/domain. Different board from the S19 — do not conflate them.
- S21 (BM1368): 108 chips per board, 12 domains, 9 chips/domain, no PIC on this generation. A count of 27 means the chain died entering domain 4 — it does not mean nine chips are dead.
- S9 (BM1387): 63 chips per board, 21 domains, 3 chips/domain.
Pull the log with dmesg | grep -i "chain\|asic\|hash" and compare the reported count against those numbers. A count landing exactly on a domain boundary points at boundary hardware — a cold solder joint on a CLK coupling resistor, or the level shifter — not at a dead ASIC. A count landing mid-domain points at a CI/BO resistor joint instead.
Spring: the second half of the problem
Dust plus repeated condensation is a conductive slurry — clean boards before the shoulder season, not after (dusting procedure). If you pull a heatsink during that clean, thermal paste must be reapplied (Fujipoly SPG-30B).
Common mistakes
- Unboxing a frozen miner in a warm shop and powering it the same hour.
- Restarting a cold, faulted machine repeatedly instead of raising intake temperature.
- Reading a cold-start fan fault as a dead fan, and replacing hardware that was never broken.
- Blaming chips for a domain-boundary break that reflowing one resistor fixes.
Related
- Optimal environmental conditions for operating ANTMINERs
- Lack of hashboard / missing hashboard
- ASIC Fault Finder — match your symptom to a repair path.
- Start a repair — board-level rework: domain-boundary reflow, BGA replacement.
