Graphics cards run hot by design and have more temperature sensors than people realise. Knowing which number you are looking at matters.
What is normal
Idle: 30 to 50 °C. Many cards stop their fans entirely below a threshold, so a warm idle with silent fans is intended behaviour.
Under load: 65 to 85 °C for the core.
Memory junction and hotspot: these read higher than the core, frequently by 15 to 25 °C, and that difference is expected rather than alarming.
The sensor that matters
Cards report several temperatures. The core temperature is the headline figure. The hotspot, or junction, is the hottest point measured anywhere on the die. Memory junction temperature is measured on the memory modules.
A large gap between core and hotspot, beyond about 25 °C, can indicate uneven contact between the die and the cooler, which is worth investigating on a card that used to show a smaller gap.
Memory junction temperature matters on cards with GDDR6X, which runs notably hot. Figures approaching the manufacturer limit under sustained load suggest airflow or thermal pad problems.
What to do about high temperatures
Check case airflow first. A card in a case with poor airflow recirculates its own exhaust.
Check for dust in the card heatsink, which is the most common cause of a gradual increase on an older card.
Check the card has room to breathe. A card pressed against a glass panel or a drive cage has restricted intake.
Consider a custom fan curve. Factory curves prioritise quiet over temperature, and a more aggressive curve trades noise for a cooler card.
Repasting a card is effective on older cards where the original paste has dried, but it usually voids the warranty. It is a reasonable step for a card out of warranty and a poor first resort for one still covered.
Undervolting, the underrated option
Modern cards are configured conservatively by the manufacturer, applying more voltage than most individual cards actually need for their rated clock speeds.
Undervolting reduces the voltage while holding the clock speed, which lowers power draw and therefore temperature and fan noise, usually with no measurable performance loss. On some cards it slightly improves sustained performance, because the card spends less time reducing clocks to stay within its power limit.
It is done in the card manufacturer tuning software, it is reversible, and it does not void a warranty the way physically opening a card does. For a card that runs hot or loud, it is the first thing to try.
Test stability afterwards with a demanding game for an extended period rather than a short benchmark. An undervolt that is slightly too aggressive shows up as a crash after twenty minutes rather than immediately.
Why a card slows down when it is hot
Cards manage themselves against several limits simultaneously: temperature, power draw, and voltage. Whichever it reaches first sets the clock speed, and most cards under sustained load are limited by power rather than temperature.
This is why lowering temperature sometimes produces less improvement than expected. If the card is holding back because it has reached its power limit, better cooling does not release it; only raising the power limit would, and that increases heat again.
Monitoring software usually shows which limit is active, often labelled as a performance cap reason. It is worth checking before spending effort on cooling that may not be the constraint.