GPU temperature and lifespan: what the physics actually says
"Cooling kills GPUs" — the forums argue about it over every cup of tea. The good news: this is not a matter of opinions. Semiconductor and solder aging is a half-century-old body of reliability engineering, written up in journals and component datasheets. Here is the digest — what reliability engineers know, and what follows for a miner running cards 24/7.
The short version (for the no-time-to-read)
- Running at 70 °C around the clock does not mean "dies tomorrow". Wear accelerates with temperature, but not the way horror stories tell it: within its spec, the chip is engineered for years of this mode.
- Steadily hot beats hot-then-cold. Solder and PCB age from temperature swings, not from high numbers as such.
- Every +10 °C roughly halves the life of vulnerable parts — that is why you watch the hottest node first (on modern cards, the memory).
- Power limits are physics, not emotion: fewer watts → less current inside the chip → slower electromigration. And less heat for the whole farm.
- Coolers and dust are consumables. A temperature creeping up at the same overclock almost always means dusty heatsinks or a dying fan.
Inside the chip: electromigration (Black's equation)
Back in the 1960s, engineer James Black described why the metal interconnects inside a chip slowly "migrate": steady current literally pushes conductor atoms around, and over time thin traces thin out until they break. His equation (Black's equation, still standard in the industry) says: the lifetime of these interconnects falls exponentially with temperature and steeply with current density.
Translated into miner language: a card with raised power and memory pushed to its ceiling runs with significantly higher current — so every compute node ages along the electromigration curve faster. Not "burning out instantly", but wearing at a rate that grows with every extra watt and degree.
Important: the chip is engineered to survive decades at nominal parameters. The point is not "a card must never die", but "a card is expected to outlast the income it brings". Pushing power shifts both curves.
Why "letting the card rest" does nothing: thermal cycling (Coffin–Manson)
Half a day of mining, then off for the night — you think the card got to rest? Failure science says the opposite: solder and the PCB age from the temperature swing per cycle. The number of cycles to failure falls, roughly, with the square of the swing — this is the Coffin–Manson law, the foundation of electronics reliability math.
The mechanics are simple: copper, plastic and silicon expand differently when heated, and every "heat up — cool down" cycle wearies the solder a little. A steady 65 °C mode goes through this physics gently — its internal swing is tiny. But "85 in the morning → 25 by the evening" day after day accumulates microcracks in the solder many times faster. Hence the practice of serious farms: cards run around the clock in even heat instead of "resting" at night.
For the same reason, do not kill the farm with a breaker "until tomorrow": every such stop is another deep heat-up/cool-down cycle of the whole board. Stops, when they are needed, should be rare and planned — not "let's just shut it off overnight, let it rest".
The hottest spot is not the one the sensor shows
On cards with modern memory, the chip's sensor can show a comfy "70 or less" while the memory itself (GDDR6X and its kin) runs noticeably hotter, close to its limits: its heat has to travel through several PCB layers and thermal pads before any sensor sees it. Miner classic: swap the cheap factory thermal pads for decent ones — and memory temperature drops by a dozen-plus degrees at the same overclock.
What follows practically:
- the numbers on screen are a guideline, not the truth: on newer cards, watch the memory first and the die second;
- thin factory thermal pads are a typical heat bottleneck; replacing them on schedule is maintenance, not "repair";
- the cooling system fails by degrees, not "suddenly": dusting the heatsinks and getting real airflow between cards buys back more temperature than any overclock tweak.
The 10 °C rule: how everything else ages
Electrolytic capacitors (motherboard, PSU) and fans age by the rule their own manufacturers use for ratings: every 10 °C above the rated operating temperature cuts the service life roughly in half — and the rule covers fan bearings too. One simple consequence: a PSU that has six cards' hot exhaust blown into it around the clock will not deliver its rated thousands of hours — the same unit with a cool intake would.
Scheduled maintenance (cleaning, replacing tired fans, listening to your power delivery) is not fuss — it is a way to push the whole farm's life cycle back by years.
Checklist: what to do with this knowledge
Keep temperature stable, not "ideal"
Even 75 °C around the clock is kinder to the solder than 60 by day and 95 at peaks. Power limits and decent airflow produce a flat curve.
Do not add current where there is no income
Watts beyond what pays raise die wear (Black's equation) — and, just as importantly, accelerate aging in the memory and power stages. Saving on heat is saving on the card.
Watch the hottest node
The bottleneck is usually memory or the power delivery, not the die. Track it first, and refresh thermal pads on schedule.
Do not switch the farm off "for the night"
Steady warm mode — fewer solder-fatigue cycles. If stops are unavoidable (expensive tariff, noise, repairs) — make them rare and planned, not "because everyone does it".
Clean on schedule
Dust on heatsinks and a struggling fan are the cheapest causes of heat. Catch them by ear and by rising temperatures, not after the fact.
Where this leads: holding the point automatically
All the rules above reduce to three jobs that have to run around the clock: watch the hottest node's temperature, keep overclocking inside the PSU budget, and cut watts the moment a card starts wobbling around it. That is exactly what OCBot does: per card it tunes overclocking inside your power limit, holds it after hours-long stability verification, and watches the numbers 24/7 — reporting a real failure in Telegram before it shows up on your revenue chart. Everything this article describes, it does on its own while you sleep.
This article is general information summarizing public reliability science (Black's equation for electromigration, Coffin–Manson for solder fatigue, and the lifetime ratings used by capacitor and fan manufacturers). Always check your specific card's spec for its exact limits. OCBot is an independent project: not affiliated with NVIDIA, HiveOS or Hiveon; their names are mentioned only to describe compatibility. Use at your own risk — see the disclaimer.