Case study · Hardware diagnosis
The card wasn’t broken. The heat path was.
An RTX 3080 throttling under sustained load. The obvious move was to replace it. The cheaper answer was to find out where the heat was actually going first.
The problem
A Gigabyte Eagle RTX 3080 could not hold a sustained compute load. Under anything continuous it would throttle back — the classic pattern of a card that is thermally limited rather than defective. Left alone, it would have kept running slow indefinitely, and it would have been easy to write off as a bad card.
The tempting diagnosis is “it runs hot, replace it.” That is an expensive guess. Card-level temperature sensors report a number, but they do not tell you where the heat is stuck, and on this generation the distinction matters: the memory and the core do not share a heat path.
How it was diagnosed
- Look at the whole card, not the sensor. A FLIR thermal camera over the card under load, to see where heat was actually concentrating rather than trusting a single reported figure.
- Verify the camera against a contact reading. A K-type thermocouple attached at the hot spot the camera identified, read on a separate meter. Thermal cameras are affected by surface emissivity — a shiny heatsink can read far cooler than it is. Two independent instruments agreeing is what turns a suspicion into a measurement.
- Work out which component owns that heat. The hot spot located the fault to the memory heat path, not the GPU core — which is what determined the fix, and ruled out the expensive conclusion.
- Change the heat path and re-measure. Correct thermal pads at the right thickness so the memory could actually transfer heat into the cooler, then the same load test again.
The result
The card went back to sustained duty and held it — running continuously rather than sawtoothing between throttle points.
The whole diagnosis was done on a live stream, so the measurement was not reconstructed afterwards from memory — you can watch the readings being taken.
What I would tell you honestly
One thing was left unresolved. I suspected the sheathing on the thermocouple wire was thick enough to affect the contact reading, and I said so at the time rather than presenting the number as final. That is the standard here: you get the measurement and its limitations. A consultant who never mentions the uncertainty in their own data is not being careful, they are being quiet.
Why this is the method
Nothing about this is specific to graphics cards. A workstation that has been slow for six months, a network that drops at the same time every afternoon, a mining rig that loses hashrate overnight — the shape of the work is identical:
- Measure before concluding, with an instrument rather than an impression
- Confirm the measurement a second way before acting on it
- Change one thing, then re-measure to prove the change did it
- Say what you are still unsure about
Parts-swapping produces a bill and no understanding. This produces a fixed machine and a reason.
Something behaving like this?
Thermal faults, intermittent throttling and machines that are “just slow” are the most common things I get called for — and the most commonly misdiagnosed.