What Nvidia actually removed
When RTX 5090 reviews went live in late January 2025, hotspot temperature vanished from monitoring tools on Blackwell cards. Nvidia had stopped exposing it through its public API: tools reading it got a stuck, meaningless value of 255, and GPU-Z 2.62.0 hid the field entirely when it saw that value, while keeping hotspot readings for RTX 40.
Asked why, Nvidia told German overclocker der8auer that API changes mean the hotspot reading is no longer accurate or relevant for gamers. So on an RTX 50 you get core temperature and memory temperature — but not the number that actually governs throttling.
The repair-shop finding
Brazilian repair specialist Paulo Gomes, working with fellow technician Sidnelson, received a Gigabyte RTX 5070 Ti sent in for overheating. In Windows, everything looked fine: average GPU temperature sat around 67–68°C. Nothing a normal owner would flag.
He booted the card into a bare-metal Linux environment and ran MODS (Modular Diagnostics Software), Nvidia's internal test suite used on factory floors and for RMA work. It is not a public download and does not run under Windows, which intercepts the low-level calls it needs. MODS read the hotspot sensor consumer tools can no longer see.
The hotspot hit 107°C almost immediately under load — Nvidia's thermal limit — and the log showed the card hitting that ceiling five times in a single test run, dropping clocks each time, cooling slightly, then climbing straight back.
Cause: bad factory paste. The TIM had squeezed out into a ring around the die's perimeter, leaving the center — the hottest silicon — nearly dry. After a cleanup and repaste, the hotspot settled at roughly 100°C and the throttling stopped.
| Fact | Detail |
|---|---|
| What changed | Hotspot dropped from Nvidia's public API, Jan 2025 |
| Nvidia's reason | "No longer accurate or relevant for gamers" |
| Sensor status | Still physically present and readable |
| Windows reading | ~67–68°C average GPU temp |
| Actual hotspot | 107°C throttle limit, hit 5x in one run |
| After repaste | ~100°C hotspot, no throttling |
What this is not
Some coverage is running with "widespread RTX 50 thermal issues." Be careful with that. Gomes says his lab has seen several Nvidia cards with the same signature — poor cooler contact or bad paste — but every one arrived because it was already misbehaving. That is a selection-biased sample by definition. It proves the sensor still exists and that a bad paste job can be invisible to the owner. It does not establish a defect rate across the RTX 50 fleet, and Nvidia has not publicly responded.
We are also not telling you to install MODS. It is not a consumer tool, and running Nvidia's internal RMA diagnostics on your own card is not a sensible weekend project.
How to catch a throttling card without the sensor
You cannot read the number, so test the behavior instead. Throttling has a signature:
- Run a sustained load, not a 30-second burst. Loop a demanding game or benchmark for 15–20 minutes with the case closed. TIM problems show up once heat soaks in.
- Watch clocks, not temperature. Log core clock in HWiNFO or Afterburner. A healthy card settles at a stable boost clock. A throttling card starts high, steps down, and stays down.
- Watch FPS drift. Frame rate that is fine for five minutes and 10–15% lower by minute fifteen, with no change in scene, is the classic tell.
- Compare against expected numbers. Use the FPS Estimator to see what your GPU and CPU combination should be producing at your resolution and settings, then check that against your logged averages. If you land meaningfully below a healthy card, something is wrong — thermals, a bottleneck, or power.
Can I see the hotspot temperature on an RTX 50 at all?
Not with normal consumer software. GPU-Z, HWiNFO and MSI Afterburner cannot show it on RTX 50 cards, because Nvidia stopped exposing it through the public API. The sensor is still on the die, but reading it takes Nvidia's internal MODS tool on bare-metal Linux.
Is 107°C dangerous for a GPU?
107°C is the limit at which the card protects itself by cutting clocks, so it is not going to melt anything — that is the safety mechanism working. The cost is performance: a card that keeps slamming into that ceiling gives you less than you paid for, and on Blackwell you get no warning it is happening.
Should I repaste my RTX 50?
Not as a first move. It voids the warranty on virtually every AIB card. Establish that you have a real problem first — sustained load test, logged clocks, real FPS versus expected — then RMA it and let the manufacturer pay for the paste.
