A rig that crashes every few days is telling you something specific, and it writes it down before it dies. Most operators skip the log, change three settings at once, and lose a week. Read the log first. It usually names the guilty part in under a minute.
Three crashes, three different causes
The driver reset is the polite crash. The screen flickers, the miner reports a hung GPU, and Windows Event Viewer shows “Display driver nvlddmkm stopped responding and has successfully recovered.” That’s a timeout detection and recovery event, a TDR. The GPU stopped answering long enough for the driver to restart it. Nine times out of ten that’s a memory clock that’s a notch too high.
The hard freeze is the rude one. The machine locks, the fans hold their last speed, and nothing moves until you hold the power button. Event Viewer logs a Kernel-Power 41 afterward, which only means the system went down without shutting down. It tells you nothing about the cause, and that’s why operators misread it as a power supply problem every time.
The sick card is the slow crash. One GPU drops 15 to 20 percent of its hash rate, climbs past 100C on memory junction, and keeps running. The rig looks fine from across the room while one card cooks itself toward an early retirement.
Read the log before you touch the clocks
Four places to look, in order.
The miner log. T-Rex and lolMiner both print the reason they restarted a card. “Watchdog” lines with a GPU index tell you which card, and the rejected share count tells you how long it’s been unhappy. A stable rig runs under 1 percent rejects. A card at 3 percent is begging.
Event Viewer, filter for the last hour, and separate TDR events from Kernel-Power events. A rig with repeated TDRs has a clock problem. A rig with nothing but Kernel-Power 41 has a power or thermal problem.
HWiNFO64, with memory junction temperature on the sensor list. On GDDR6X cards that number runs hot by design, but sustained 100C and above is the throttle wall, and past 105C you’re buying a new card. Log it for an hour instead of spot-checking.
nvidia-smi -q, which prints throttle reasons per card. “SW Power Cap” means the power limit is doing its job. “HW Thermal Slowdown” means the cooler is losing.
The settings that stop most of it
Once the log names the failure, the fix is usually boring.
For TDRs: drop the memory clock 100 to 200 MHz and watch rejects for a day. If the crash returns, drop again. The winning clock for a card is the highest one that survives a week, not the one that benchmarked clean for ten minutes.
For freezes under load: check what powers the risers. SATA-to-Molex adapters on USB risers have melted more rigs than any overclock. Run each riser off a dedicated PCIe cable from the PSU. Then look at transients. A card that spikes toward twice its rated draw can trip a supply that shows plenty of headroom in steady state, which is exactly why overclocking guides keep telling you to leave 20 percent on the PSU.
For the sick card: clean the fin stack, check the fan for bearing tick, and look at the paste age. Anything past 18 months of continuous service is a repaste candidate.
For random instability after a driver update: roll back. Keep one known-good driver version archived and stop letting the rig auto-update. A mining rig is infrastructure. The HWiNFO and driver versions that worked last month are the versions that should work this month.
What a stable rig looks like on paper
Hash rate flat within 1 percent over a week. Rejects under 1 percent. Memory junction 85 to 95C at the warmest card. No TDR events in Event Viewer for thirty days. Fan speeds steady instead of creeping.
Write those five numbers down somewhere and check them weekly. When one drifts, you’ll catch the problem in an afternoon instead of discovering it as a crash.