Goldshell KD5 / KD6 – Overheating / Thermal Shutdown

Cold-boot and baseline temps before touching anything. Power off, pull AC, wait 30 seconds, plug back in. Open the web UI at http://<miner-ip> with admin / 123456789 and let the miner run 15 minutes. Note chip temp per hashboard, ambient, fan RPM, hashrate. A healthy KD6 at 22-25 C ambient should sit at chip 70-85 C, fans 4,500-5,500 RPM, hashrate within 5% of 26.3 TH/s. If chip temp climbs >1 C/min from cold start you have a real thermal event, not a stale reading.

Vacuum the intake grille and fin stack from the outside. Power off the miner. Use a vacuum cleaner with a soft brush attachment to lift loose dust off the front intake grille and as much of the visible fin face as the nozzle can reach. Do not stick the nozzle into the fan blades. This step alone resolves about 30% of KD-series thermal complaints from operators who haven't cleaned in 6-12 months. Cost: zero. Time: 5 minutes per grille face.

Open the side panels and inspect for visible dust mat or fibre between the heatsink fins. Remove the side cover screws (Phillips #2 or Torx T8/T10 depending on revision). Look down through the fan grille at the fin stack with a flashlight. If you see a grey woven mat between fins, or you can run a fingertip across the fin tops and lift fibre, you are looking at packed-fin obstruction — the dominant root cause. If fins look clean, skip ahead to Step 7 (fan RPM check).

Improve room airflow / exhaust / ambient. Check intake air 5 cm in front of the front fan grille with a digital thermometer or IR gun. Goldshell's hard cap is 35 C ambient and you want 7+ C of headroom — target intake under 28 C. If the room is over 30 C: open a window, run an exhaust fan, duct the miner exhaust outdoors, or move the miner to a cooler wall. A KD-MAX dumps ~3,360 W; in a 15 m^2 closed room that heats ambient 5-8 C in the first hour with no exhaust path. Re-test thermals after 2 h of better airflow before doing anything else.

Reset firmware to factory defaults and clear any stale OC profile. In the web UI go to Settings -> Restore Factory Settings (or hold the RST pinhole for 5-10 s if web UI is unreachable). Re-pair to your network over Ethernet, log in, change the default 123456789 password, and re-add your pool. Stale or aggressive overclock profiles can hold higher voltage / frequency than the silicon wants and push chip temp 3-8 C higher than stock. Stock profile gives you a clean baseline before you touch anything mechanical.

Pull the chassis covers and deep-clean the fin stack with compressed air plus brush. Power off, AC unplugged, miner on its side. Blast compressed air (canned or shop compressor capped at 30 PSI) down the fin valleys at a 30-45 degree angle from the EXHAUST face — against natural airflow direction. Simultaneously vacuum the intake face to capture dislodged debris. Do both sides. Use a soft anti-static brush for stuck mats. Do not bend fins. Five minutes per side, repeat once. Goldshell's fin pitch is roughly 2 mm and traps lint at a brutal rate; this is the only technique that actually clears packed valleys.

Verify both fans run at nameplate RPM under load; replace any failing fan. Boot the miner, watch fan RPM in the web UI as it ramps from idle to full hashrate. Both should climb to 4,500-5,500 RPM and stay there (KD-MAX: all four fans). If one fan reads 0 RPM, stays well below the other, or audibly grinds / ticks, the bearing is gone — replace it. Stock fans are 120x120x38 mm, 12 V, 4-wire PWM, ~5,500 RPM. Available from D-Central or Amazon. Replace one and the other usually follows within 6-12 months — bench tip: swap both at once if the miner is past 2 years.

Measure intake / exhaust temps with an IR gun; document the delta. Run the miner under load for 5 minutes after Steps 1-7. With a Fluke 62 MAX or equivalent, read intake air, then exhaust air at the fin tips on the back face. Healthy KD6 delta: ~25-35 C between intake and exhaust. If delta is below 15 C, the miner isn't moving heat (clogged fins or weak fan). If delta is above 40 C, airflow is good but chips are running hot — TIM or chip-level issue. Document the numbers; you'll need them again at Step 15.

Re-seat hashboard data + power cables; check for backed-out connectors. Power off, AC unplugged. Open the chassis. Check that each hashboard's data ribbon and power connector are fully seated. Vibration over time can back connectors out 1-2 mm — enough to introduce intermittent contact, occasional sensor bogus readings, and (rarely) phantom thermal trips when one board's reading goes wild. Press each connector firmly home. Verify no bent pins, no scorched plastic, no loose strands. Re-seat; do not over-flex the ribbon.

Cross-flash the latest stable Goldshell firmware over Ethernet only (never WiFi). Goldshell's KB explicitly warns that WiFi firmware upgrades corrupt the filesystem in a non-trivial percentage of attempts. Plug Ethernet straight to your router. Download the latest stable .bin for your model from the Goldshell firmware page. In the web UI, Settings -> Firmware Upgrade -> upload. Do not power-cycle during the flash. Older firmware versions occasionally had thermal-protect bugs that triggered phantom trips; staying current eliminates that variable before you start chasing hardware.

Pull suspect hashboard; remove old thermal paste; re-paste with Arctic MX-6 or Kryonaut. Power off, AC unplugged, ESD strap on. Remove the hashboard's mounting screws (Torx T8/T10). Carefully separate the heatsink from the ASIC carrier — old TIM may bond them; gentle twist motion, do not pry. Scrub all old paste off both surfaces with 99% IPA + lint-free wipes until both are mirror-clean. Apply a rice-grain-sized dot of Arctic MX-6 (or Kryonaut, or Honeywell PTM7950 phase-change pad) to each ASIC die. Re-mount heatsink; re-torque clamp screws to spec (snug-plus-quarter-turn, do not crank).

Inspect heatsink clamp tension; re-torque per Goldshell spec. Goldshell uses spring-loaded standoffs to clamp the heatsink to each hashboard at a controlled pressure. After 2-3 years of thermal cycling, those springs can lose preload by 10-20%. Visually compare clamp height across all chips; any clamp that sits notably higher than its neighbours has lost preload. Re-torque all clamps to a consistent feel — uniform across the board. Uneven clamp pressure produces uneven chip temps within a single board; that hot chip drags the board's reported temp up regardless of fan or TIM state.

Verify thermal-pad coverage on PCH / VRM regions; replace if degraded. While the heatsink is off, inspect the secondary thermal pads between heatsink and the PCH / VRM ICs (not the main ASICs, the smaller power-control chips). Healthy = pliable, smooth, fully covering. Degraded = cracked, hardened, gaps. Replace with 1.0 mm or 1.5 mm thermal pad rated 8.5 W/m·K minimum (Gelid Extreme, Thermalright Odyssey). Wrong pad thickness is worse than no pad — too thin and you have no contact, too thick and you over-clamp the ASICs. Match the thickness of the pad you removed.

Cross-check temp sensor reading against IR gun; swap sensor cable if available. Run the miner under load 5 minutes. IR-read the heatsink fin tips above each hashboard's reported temp sensor location. If firmware claims a board is at 100 C but the IR says 50-55 C exhaust fin temp, the temp sensor is lying. On KD-MAX you can swap the temp-sensor cable header to a different board's port; if the high reading follows the cable -> bad cable. If it stays with the board -> on-board sensor. Bad sensor = Tier 4 (component-level repair, not bench-DIY). Bad cable = replace the cable, problem solved for under $5.

Burn-in 4 hours after re-paste; verify chip temp stable below 90 C. Re-assemble the miner, power on, point at your normal pool. Watch chip temps for the full 4 hours. A successful re-paste should hold chip temp 5-15 C below the pre-fix baseline at the same ambient. Hashrate should return to within 5% of nameplate. If chip temps are still climbing past 90 C with healthy fans, clean fins, and fresh TIM, the failure is silicon (chip drift) or sensor — Steps 14 and 16-18. Document burn-in numbers; you'll send them with the unit if Tier 4 is needed.

If chip temp still trips post-repaste, individual chip failure is likely — ship to D-Central. Symptoms: one specific chip position runs visibly hotter than the rest under IR, or one hashboard alone trips temp protect even after cleaning + repasting + verified-healthy fans. This is electromigration-driven silicon failure, accelerated by a history of running above 95 C. The fix is chip-level rework — desolder failed chip, replace from a tested salvage donor, reflow at 245 C peak. This is bench work with reflow gear, not a kitchen-table repair. Pack the miner safely and ship to D-Central.

D-Central bench replaces failed chips, sensors, control board components as needed. We run IR thermal cam diagnostics on every Goldshell repair to identify failed chip positions before any rework. Failed chips get desoldered with hot-air rework, replaced from our salvage donor stock (KD5 / KD6 / KD-MAX boards), and reflowed. Failed temp sensors get replaced as discrete components. Control-board issues (the SoC that runs the miner daemon) are rarer but occur; we have known-good board swaps for the entire KD lineup. No 6-week China round-trip via Zeus, no shipping a KD-MAX across an ocean — Canadian bench, documented workflow.

D-Central runs a 24-hour burn-in at nameplate hashrate before return ship. Every Goldshell repair gets a 24 h burn-in at the customer's preferred hashrate (or nameplate if not specified) before it leaves the bench. We log chip temps, fan RPM, hashrate stability, and pool acceptance ratio over the full 24 h. Anything that fails burn-in goes back on the bench. You get a one-page report with the unit when it ships back — pre-repair vs post-repair temps, what was replaced, recommended ambient and maintenance interval to keep it in spec going forward. Typical turnaround for a Goldshell thermal repair is 5-10 business days from receipt.