NerdOctaxe – Uneven Hashrate Across 8 Chips

Cold power-cycle for 30 seconds at the XT60 (not a soft restart from the web UI - the BM1370 chain does not always re-enumerate cleanly from a firmware-side reset). Wait a full half-minute unplugged, plug back in, let the miner stabilize for 2 minutes, then read the per-chip view again. Roughly 15% of uneven-hashrate reports clear on a real cold-cycle.

Revert to stock NerdOS profile: 600 MHz global frequency, 1150 mV VCORE, ECO disabled. Let it run 10 minutes at steady state. If spread closes to under 0.15 TH/s per chip, your previous tuning profile was beyond the weakest die's silicon ceiling - rebuild the OC slower in Tier 2. If spread stays wide at stock, the problem is not tuning - move to thermal and power checks.

Visually verify both AXP90-X53 cooler fans spinning at full RPM. Open the enclosure, power on, watch both fans from the side. A failed fan or one at half-RPM is the #1 misdiagnosed cause of uneven cluster hashrate - the chips are not weak, they are hot. If a fan is slow or stopped, that is your Tier 2 fix.

Check ambient temp at the miner intake with an IR thermometer. NerdOctaxe wants ambient <= 28 degC for clean stock, <= 22 degC if you are tuning to 650 MHz+. A floor fan aimed at the intake for 10 minutes is a free diagnostic; if the spread narrows with forced airflow, the room is too warm for the current tune.

Re-read the D-Central NerdOctaxe Gamma setup guide at d-central.tech/manuals/nerdoctaxe-gamma-setup-guide/ for the current firmware version's tuning ceilings and per-chip offset syntax. NerdOS updates occasionally change recommended defaults, and a stale cheat-sheet is a common cause of profiles that worked last month and are uneven this month.

Measure wall draw with a Kill-A-Watt and 12V input at the XT60 with a multimeter on DC under load. Expect ~208 W at the wall at stock 600 MHz; expect >= 11.8 V sustained at the XT60 input. Below either threshold = PSU undersized or tired. The 18-20 A minimum on the D-Central spec is not a suggestion; a bundle 15 A PSU will sag under an 8-chip demand and your uneven chips are actually uniform VCORE droop.

Swap to a known-good 12V / 20A+ / 240W+ supply. Community standard: HP DPS-1200FB or Dell 1100W server PSU with an XT60 breakout board; RC/drone 12V supplies; Mean Well LRS-350-12 industrial. Retest per-chip hashrate after 10 minutes at steady state. If the spread closes after the swap, you had PSU sag - not silicon lottery.

Re-torque both AXP90-X53 cooler mounting screws. Four M3 screws per cooler, eight total. Finger-tight plus a quarter turn - do not over-tighten or you will crack the PCB around the mounting through-holes. Loose mounts after shipping are one of the most common root causes of per-cluster thermal imbalance in the first month of ownership.

Swap the slower or silent cooler fan. AXP90-X53 ships with a 92mm x 14mm slim fan. If one is visibly slower or dead, replace with an equivalent: Noctua NF-A9x14 for quiet-but-competent, Arctic P9 Slim PWM for budget. Verify with a tach or visual comparison to the good side before reassembly.

Rebuild OC from 600 MHz stock in 25 MHz steps. At each step, run 10 minutes and read per-chip hashrate. Watch for the weakest chip to either lose hashrate or start throwing invalid-nonce flags in the serial log. Stop one step before that happens. That is the silicon-lottery ceiling for this specific NerdOctaxe - no two are identical. Save the profile.

Re-paste both AXP90-X53 coolers. Power off, pull both coolers (four M3 screws each, disconnect fan headers), inspect factory thermal paste. If dried, unevenly spread, or showing pump-out voids, clean with 99% isopropyl alcohol and lint-free wipes. Apply Arctic MX-6 or Thermal Grizzly Kryonaut in a uniform thin layer (grain-of-rice center + small corner dots for an 8-chip spread), remount, retorque. Use thermal pad on any TPS546 / PMIC contact point if the original layout used one.

Verify per-chip enumeration via UART serial console. USB-C to laptop, 115200 8N1 in PuTTY, screen, or minicom. Reboot and watch asic_enumerate output. Healthy: `8 chips detected`. Faulted: anything less, or retry loops on specific positions (especially #6 or #7 on early revs). A chain that enumerates 7 of 8 and reports the missing chip as weak on the dashboard is not weak - it is gone. That is a reflow or repair call.

Use NerdOS per-chip frequency offset. Newer NerdOS builds expose per-chip frequency offset in the web UI or config file - set the weakest chip -25 or -50 MHz below the global target so the strong chips are not handcuffed. This is the open-source-firmware moat: impossible on Bitmain stock, painful on LuxOS, trivial on NerdOS. Verify per-chip hashrate converges toward uniform after the offset.

Reflow the worst-performing chip. If one BM1370 position consistently reports low after thermal, power, and tuning are all clean, it is a solder-joint / BGA issue. Remove hashboard from chassis, flux the target chip's BGA perimeter, preheat bottom of board to ~150 degC, top-side hot air at 310-330 degC for 30 seconds, cool naturally, re-apply thermal paste, reassemble. Low risk on BM1370 packaging; high reward if the joint was the actual failure.

Downgrade NerdOS one version via the D-Central web-flasher. If a recent release introduced a chain-enumeration or VCORE calibration regression, rolling back one build is a 10-minute diagnostic. Keep the known-good .bin backed up locally; a web-flasher that loses network mid-flash is how people end up with bricked ESP32-S3 boards.

Stop DIY and book a D-Central repair when: you have reflowed a chip and the same chip fails within 30 days; two different chip positions on the same board consistently under-hash after thermal and power are clean; or you see visible damage, burn marks, cracked ceramics, or smell cooked components. The NerdOctaxe is hand-assembled in Laval, Quebec - the same bench that built it will fix it.

What D-Central does at the bench: per-chip isolation with a programmable 12V test load and logic-analyzer probing of the serial enumeration chain; salvaged-grade or new-old-stock BM1370 chip swap (we pull chips from the same Antminer S21 Pro silicon family we repair every week); full hashboard reflow and reseal; 24-hour burn-in at nameplate 600 MHz before the unit ships back. Turnaround 5-10 business days Canada-wide.

Ship safely: anti-static bag for the hashboard or the whole assembled unit, double-box with 5 cm of foam on every side, include a note with observed per-chip hashrate numbers, firmware version, PSU model, and ambient temp range. The more context D-Central gets up front, the faster the bench ticket closes and the cheaper your repair bill lands.