Throttling: why a GPU drops its own clocks
The card seems alive: the miner runs, shares flow — but the clocks sag and the hashrate floats. This is not a breakdown and not a "bad driver": the card drops its own clocks to escape overheating or power overload. That is throttling, and it is a useful signal if you know how to read it.
The short version (for the no-time-to-read)
- Throttling is not a failure, it is self-protection. The card lowers clocks to survive; the question is not "how to disable it" but "what it is protecting against".
- Two main kinds: thermal and power. The first means "too hot", the second — "power budget exceeded". They are fixed in opposite ways.
- The monitoring signature: core clock falls as temperature rises or as the card pins against its power limit, and hashrate floats along.
- Memory throttling exists separately — GDDR6X has its own limit and its own temperature sensor.
- A voluntary minus 10–15% of watts almost always beats the same minus 10–15% the card will bite off by itself at the worst moment.
Thermal throttling: the card is hot
Every chip has a temperature past which the controller starts dropping clocks to stay below it. The mechanics are honest: the hotter it gets, the lower the clocks, the less heat. The card finds equilibrium "on the edge" — and sits there, losing performance.
Memory is a separate story. On GDDR6X cards the memory chips run hotter than the die and have their own throttling: the core is still fine while memory already sheds its effective speed (details in the hot memory article). So "core at 65 °C, all good" does not cancel quiet memory throttling next door.
The causes of overheating are almost always mundane: dust in the heatsink (the maintenance schedule), dried-out thermal paste, a hot room in summer (on heat), a suffocating case or a box without exhaust.
Power throttling: not enough watts
The second kind has nothing to do with temperature: the card has a power limit and cannot exceed it. If the load and the overclock demand more than allowed, the card lowers clocks to fit the budget. In monitoring this shows as a flat pin against the power limit on a cool card.
An important nuance: pinning against the limit is not always bad. On many algorithms a card sitting steadily on its power limit runs more efficiently than the same card unlimited at maximum clocks. The question is whether you chose that limit deliberately — or it sits "factory default" and gets in the way.
Telling the kinds apart is easy: watch the temperature and power graphs together. Clocks falling with rising temperature — thermal. Clocks low on a cool card pinned at the watt limit — power.
What to do in each case
- Thermal: cooling first — blow-out, intake/exhaust, thermal interfaces. Not enough — cut watts: less heat at the same fan speed.
- Power: if the goal is hashrate and temperatures allow, raise the limit within reason; if the goal is efficiency, keep the pin but pick clocks for it deliberately.
- Memory throttling: lower the memory overclock — it is the main source of GDDR6X heat; the lost memory hashrate often returns as stability.
- A floating hashrate, "fine then sagged" with no visible heat — check for periodic throttling: that is exactly what a preset sitting just past the edge looks like.
Where automation fits
Throttling is the card searching for its own working point — crudely, at the last moment. OCBot does the same search in advance and gently: it tunes an overclock where the card pins against neither temperature nor the power limit, and verifies stability over long runs. The result is an even hashrate without sagging clocks and midnight surprises.
This article is general information summarizing public practice; actual limits and behavior depend on your card model. OCBot is an independent project: not affiliated with NVIDIA, HiveOS or Hiveon. Use at your own risk — see the disclaimer.