GPU Thermal Throttling: How to Diagnose and Fix It

Three different throttles

A graphics card reduces clocks for several independent reasons, and each has a different fix. Treating them as one problem is why the obvious remedies so often fail.

The second row is the one that surprises people. On cards using high-speed memory, the memory junction sensor commonly runs far hotter than the core — it is normal for a gap of 20 to 30 degrees to exist between them. Someone checking only core temperature sees a healthy number and concludes GPU thermal throttling is not the problem, while the memory is the component actually forcing clocks down.

The third row is worth internalizing too: a card sitting at its power limit is not faulty. NVIDIA specifies a total graphics power of 360 W for the RTX 5080 and 450 W for the RTX 4090, and reaching that figure under load is the card working as designed.

How to diagnose it properly

Log rather than glance, and log during sustained load rather than a benchmark run.

1. Choose a real sustained workload — twenty minutes of the game or render you actually care about, not a thirty-second synthetic test.
2. Log five values over time: core temperature, memory junction temperature, core clock, board power draw, and fan speed.
3. Find the moment clocks first drop, then read what the other four values were doing at that instant. That correlation is your diagnosis.
4. Note the ambient room temperature. A card is only ever as cool as the air it is given, and a 10-degree warmer room moves everything by roughly 10 degrees.
5. Check whether fans reached maximum. If clocks dropped while fans were still at 60%, the fan curve is the problem rather than the cooler.

That last check catches an unreasonable share of cases. Factory fan curves are tuned for quiet operation, and many cards will happily throttle rather than spin up.

Fixes in order of cost

• Clean the heatsink and fans. Dust in the fin stack is the most common cause of a card that used to be fine. Free, and frequently sufficient.
• Improve case airflow. A graphics card recirculating its own exhaust cannot be cooled by any fan curve. Ensure intake at the front and exhaust at the rear or top.
• Raise the fan curve. Accepting more noise above 70 °C is the cheapest performance recovery available.
• Remove obstructions. A card mounted directly against a drive cage or a glass panel with 15 mm of clearance is starved regardless of specification.
• Lower the ambient temperature. A closed desk cabinet can add 10 degrees to everything.
• Undervolt or cap the power limit. Costs a little peak performance, returns a lot of thermal headroom.
• Replace thermal pads or paste. Effective and invasive; usually voids warranty.

Power limits and undervolting

These two approaches sound similar and behave differently. Capping the power limit tells the card to draw less, and it responds by lowering clocks — simple, safe, and it costs performance proportionally. Undervolting asks the card to hold a given clock at a lower voltage, which reduces heat while keeping performance, and it requires testing for stability because not every card tolerates the same curve.

The economics favor both. Graphics cards are pushed close to the top of their efficiency curve from the factory, so the last few percent of performance costs a disproportionate share of the power. Capping power to 85 or 90% typically costs low single-digit performance and drops temperatures substantially — which is why it is the most reliable single fix for GPU thermal throttling on a card that is otherwise healthy.

Test stability properly after undervolting: an unstable undervolt presents as a crash hours into a session rather than immediately.

When it is the thermal interface

Suspect the interface material when a card that was previously fine has degraded gradually over years, when the core-to-hotspot gap has widened compared with earlier logs, or when memory junction temperatures are extreme while the core is fine — that pattern points at memory thermal pads specifically rather than the main paste.

Two cautions. This voids the warranty on most cards, so exhaust the free fixes first and check whether the card is still covered. And pad thickness matters: fitting pads that are too thin leaves an air gap and too thick lifts the heatsink off the core, either of which makes the situation worse. Measure or source the manufacturer’s specification rather than guessing.

The checklist

• Log memory junction temperature, not just core temperature.
• Use a twenty-minute real workload, not a short synthetic run.
• Check whether fans hit maximum before blaming the cooler.
• Clean the fins and fix case airflow before spending anything.
• Cap power to around 90% as the highest-value single adjustment.
• Treat repasting as a last resort, and check warranty first.

Worked in that order, GPU thermal throttling is usually solved for free within an hour — and the diagnosis, once you are reading the right sensor, is normally obvious in the first log. Our cooling and power guide covers building a case that does not create the problem in the first place.