Defective Heat Sinks Causing Garbage Gaming

<i>&gt;&quot;When I told the customer of my suspicions they said that they had recently replaced their processor heat sink. When they put the old heat sink back on their performance problems went away.&quot;</i><p>A person who buys and installs their own heatsink, but can&#x27;t troubleshoot a temperature-driven performance problem impresses me as an odd combination.<p>I expect providing support to members of the &quot;enthusiast&quot; PC market is an interesting challenge.

Cleaning the CPU heat sink and fan from dust (and in my case, cat hair) can sometimes give a surprising performance boost. Especially on cheap laptops that are often sold with inadequate cooling.<p>CPU thermal throttling can have a huge effect on performance.

Its common in hardware enthusiast forums and discussion now when there is a performance problem for there to be 2 initial questions. The first is regards to clock speed and the second is temperature. Its really common for clock throttling to be the underlying cause of performance problems on today&#x27;s systems and temperature is also the number one cause.<p>But the second major cause, representing about 30% of cases is software&#x2F;firmware bugs in the drivers for the graphics card or the Windows scheduler doing odd things for the CPU under a particular workload. As more of the performance of components becomes dependent on boosting and clock increasing in order to save power the more we are seeing inconsistent behaviour and problems with the implementations.

&gt; Thermal throttling is extremely difficult to detect in xperf traces. It is done automatically by the CPU or the motherboard, and the operating system (OS) doesn’t realize that it is happening. I use the xperf toolset for these investigations but its CPU frequency graphs show the CPU running at full throttle and its power management events say that all is well, and yet…<p>At least with a Sandy Bridge CPU on Linux you get power management events which show up in dmesg and the reduced frequency is correctly reported. Definitely a case of the drivers&#x2F;monitoring software not looking for the right things rather than the hardware not exposing this information.

This should be made more visible by OS vendors.<p>Windows popping up one of those alert baloons would be very helpful.

I recently discovered &#x27;turbostat&#x27; for reading actual CPU frequency on Linux: <a href="http://lxr.free-electrons.com/source/tools/power/x86/turbostat/turbostat.c" rel="nofollow">http:&#x2F;&#x2F;lxr.free-electrons.com&#x2F;source&#x2F;tools&#x2F;power&#x2F;x86&#x2F;turbost...</a><p>Bruce&#x27;s direct testing approach has a certain elegance, but instead of timing a loop &#x27;turbostat&#x27; queries the CPU&#x27;s performance registers. So in addition to being handy if you are running Linux, it could serve as a guide for developers of other systems trying to figure out which MSR&#x27;s are needed and how to interpret them.