Thermal throttling

From Prime-Wiki
Jump to: navigation, search

(Originally written by Ethan Hansen on Mersenneforum)

Your reported iteration speed of 113ms (I'm assuming a single instance of Prime95 running here) is ~25% of what one would expect for your CPU. That is suspiciously close to a default throttling mode.

Download the ThrottleWatch utility from Panopsys. It monitors actual operating frequency - not the value reported by the O/S - and flags if the CPU issues a PROCHOT# signal. If you see throttling, either your thermal interface material (grease, pad, etc.) between the die and heatsink is not evenly applied or the heatsink surface is not flat. I am not familiar with the Arctic Cooling product, so I don't know if it is prone to quality control problems.

An on-die thermal sensor value of 35C is improbably low for a CPU maxed out by Prime95. This points again to throttling. The CPU has independent sensors for the on-die temperature diode that applications can read and the thermal monitor used for PROCHOT# and THERMTRIP#. These sensors are located relatively far from each other, and having one be hot enough to cause a problem (throttling) while the other reports a low value (35C) is not unheard of. Prime95 is particularly bad in this regard, as sections of the code hammer certain functional blocks particularly hard. A potential check for this is to go into the advanced menu in Prime95, select the affinity option, and set Prime95 to operate only on CPU0. Monitor the temperature until it stabilizes, then stop Prime95, set the affinity to CPU3 (assuming HT is enabled; CPU1 if not) and run it again. If there is a difference in reported CPU temperature, throttling is almost certainly the problem.

(For versions 25 and above, the affinity option is found in Test -> Worker windows)

Random notes:

  • Some motherboards require that you explicitly enable ACPI thermal monitoring for the function to work properly.
  • The 9xx series has an ACPI mode called "On-Demand" that forces the processor to run at a specified duty cycle ranging from 12.5% to 87.5%. Look through your BIOS setup to make sure that this feature is not enabled (settings usually found amongst the other power saving features).
  • Do you have the fan speed reduction (TCONTROL) option enabled in your BIOS (assuming it is present at all)? This allows the system to reduce fan speed as long as the CPU temperature is below a specified value. As noted above, the physical separation between the thermal sensors used for the on-die monitor and the thermal trip circuitry can cause a very few throttling cycles when a CPU intensive application first loads if TCONTROL is used. This can be exacerbated if your fan operates similarly to the OEM version, where fan speed is also controlled by a thermistor in the fan hub.

External Links