Introduction

With the continuous advances of PC technology, test benches must be upgraded to keep up with the times. This is especially true for the testing of graphics cards, and while the need to constantly update and innovate is not as paramount for CPU coolers, the setup still needs to be viable and offer tangible results. We previously used an Intel Core i7-8700K, which is still a decent platform as most systems being built are likely to be based around the more affordable Intel and AMD processors. That said, Intel's processors continue to see higher and higher TDPs, and AMD Ryzen continues to gain in popularity. As such, it's time for a massive update not only to our testing platform, but also the testing methodologies.



AMD's rise to prominence is no secret, and while we will be testing on an AMD platform, we have not abandoned Intel. The decision was made to use two test systems rather than one for a better look at the performance CPU coolers can offer on both mainstream platforms. In this article, we will thus discuss the used systems, updates to the testing suite, and changes made to our testing methods.

Testing Systems, Platforms and Settings

For the AMD test platform, we opted for a Ryzen 9 3900X. It will be featured in the photos and as the primary focus during the installation procedure since AM4 is the more popular platform currently. To find our starting point on the AMD system, we used an AMD Wraith Prism cooler with the fan set to maximum performance via the tiny switch. With stock BIOS settings, we then used Blender 3D to determine our processor's default all-core clock speed, which was 3.85 GHz. This is where our processor settled during our testing with the AMD stock cooler. Thus, we manually set the CPU all-core clock to 3.85 GHz and left all voltages and settings alone, which removed the issue of the clock speed ramping up on its own based on temperature headroom, which would screw with the cooler test results. While not an ideal situation, it's the only way to get an apples-to-apples comparison.

For overclock testing, the CPU was pushed to 4.0 GHz with the voltage bumped from the stock 1.099 V in Blender to 1.22 V for overclock testing. We also had to increase the SOC voltage to 1.125 V to maintain stability, which had Ryzen Master report an increase from 105 W on the CPU and 39 W on the SOC to 150 W on the CPU and 44 W on the SOC. Meanwhile, it should be noted that while software voltages are not as reliable as hardwired readings, it is done here for reference purposes. Testing power draw at the wall, we noted a 50–60 W increase from stock to overclock.


On the Intel test platform, we opted for an Intel Core i9-10900K. To find our starting point, we used a Noctua NH-U12S with the fan set to maximum performance. At stock BIOS settings, we disabled all multi-core enhancement features and stuck with Intel's specifications. We then used Blender 3D to determine our processor's default all-core clock speed after the initial boost period ended, which was 4.3 GHz. To ensure testing remains fair on all coolers, we then manually set voltage to Vcore at the time of testing, which was 1.01 V, and limited the CPU to the predetermined clock speed. Again, while not an ideal situation, it is the only way to get an apples-to-apples comparison between coolers that is consistent.

For overclocked testing, the motherboard has multi-core enhancement enabled and all limits removed. We then set the CPU to 4.8 GHz at 1.21 volts. This results in the CPU going from the 125 W limit defined by Intel all the way up to roughly 200 W when placed under sustained load. While this isn't the peak power this CPU can draw when running multi-core enhancement, it is a sustainable overclock for long-term testing. Testing power draw at the wall, we noted a 70–90 W increase from the stock settings compared to the overclock settings, with the CPU averaging 194–205 W under load.

Temperature Monitoring and Limits

We have completely changed how we monitor temperatures. We opted for AMD's Ryzen Master on the AM4 platform, with the package temperature limit being 95°C. We found Ryzen Master to be accurate, easy to use, and quite robust, and we get a quick look at both CPU and SOC power information at a glance. Again, while the software is not a true replacement for hardware-level monitoring, it still allows for a good look at typical power usage. I will note that the CPU can go beyond the 95°C; however, this is our cutoff point. To prevent degradation of the processor, the system is shut down if temperatures significantly exceed the cutoff point.

For temperature readings on Intel's Z490 platform, we continue to use AIDA64 with the Tj. Max (temperature junction maximum) set to 100°C. While the Tj. Max can be raised on the Intel Core i9-10900K depending on the motherboard, we stuck with the default as not all boards may offer that setting, and we would again prefer long-term CPU longevity for testing.

Fan Noise and RPM Readings

When testing CPU coolers, the rest of our system is completely passive—no fans other than those on the CPU cooler are running. This includes the graphics card and the power supply and allows for testing the noise levels of the CPU cooler only since all other contributors of noise have been removed. This was accomplished by disconnecting the fans on the graphics card and utilizing a semi-passive power supply that runs fanless at low to moderate load levels, which means the fan does not ramp up even if the CPU is heavily stressed.


Note the chart above is for reference only. The coolers shown had their noise levels taken at 6 in. / 15 cm. However, stepping away three feet results in a sizable drop in perceived noise. This was done to accommodate the new testing area, where our noise floor is a bit higher than it was previously. We also expanded noise testing to include 25%, 50%, 75%, and 100% PWM settings. For example, a cooler like the Noctua NH-D15 has a noise profile of 34, 40, 43, and 49 dBA. While I won't go into explaining how a logarithmic scale works, the simple fact is that noise levels will appear higher in the charts going forward, compared to the old data. This shift upwards is a byproduct of both the reduced distance and having now replaced the very old Pyle SPL meter with a BAFX 3370 sound level meter.

It should be noted that different meters will give different results as well. This is why decibel readings between different review sources should not be compared. Plenty of variables are not accounted for—instead of looking at this and previous reviews, it's better to compare where the coolers land in relation to each other to better grasp the noise level you can expect.

Yes, this means even the results from our older reviews are not comparable to new reviews in any way, shape, or form. This cannot be stressed enough. That having been said, any cooler with a measured sound level below 45 dBA in our tests is very quiet. Anything below 50 dBA can be considered relatively unobtrusive for the majority of end-user situations unless the cooler is directly next to you (e.g., 1 ft / 30 cm or, in our case, 6 in. / 15 cm).

Noise normalized testing will now make an appearance in cooler reviews. All CPU coolers will be tested at a set 45 dBA, which, as noted above, is very quiet owing to how we test. At this noise level, a cooler can be considered unobtrusive at worst or nearly silent at best for the majority of users, depending on your chassis of choice. Also note that some coolers do end up quieter than our stated noise normalized settings. However, the majority end up equal to or louder at maximum RPM, with more than a few getting very close to 60 dBA.

Speaking of RPM, we also record the cooler's fan speeds at the same 25%, 50%, 75%, and 100% PWM settings to give users a direct reference point from which the dBA readings were achieved.

Temperature Testing

Idle temperature testing consists of booting the system and loading the temperature monitoring software, at which point the CPU is left to sit idle for 15 minutes before we record the average idle temperature. The system is not connected to the Internet, so our results are not influenced by potential driver or Windows updates and miscellaneous background applications.

The primary load test is done with Blender. Using the BMW test scene, we loop this test for 15 minutes. While there is a slight delay between loops due to the length of the test, the results are consistent and represent a real-world workload. The peak temperature noted during testing is what is carried over. This is because the peak temperature is what will trigger throttling, instability, etc. Sure, a cooler can toe the line with an average temperature of say 95°C, but if it has the CPU peak at 100°C, for example, you are likely to see thermal throttling depending on the processor, BIOS configuration, etc. A situation like that also changes the processor's overall performance, and we thus do not use the average temperature. For AMD, if a CPU cooler spikes beyond the peak 95°C package limit on our Ryzen 9 3900X, the system is shut down to avoid potential processor degradation. Considering our testing includes entry-level to high-end coolers, mitigation of said potential damage is important. To check for degradation, we will be periodically retesting the Deepcool Assassin III and Noctua NH-U12A for perceptible changes.

The final test is AIDA64. We use the software's Stability Benchmark to hammer the CPU's floating-point unit. While this workload is not "realistic," it separates average from top-tier offerings—it also serves as a good test for those who may be in warmer climates because if the CPU cooler can handle this test, it has some headroom to spare. On our AMD test bench, this test shows a sizable increase in temperatures. On Intel with the Intel Core i9-10900K, the difference is much smaller.

All the above tests are repeated three times in total. The entire process is then restarted for noise normalized testing. For those wondering, this works out to 18 hours of testing excluding the time taken by reboots and swapping from one system to another. That figure also doesn't include all the other work necessary to produce a review, so if you want to ask why we do not test with standardized fans, etc., this is it. There is also the fact that the vast majority of users tend to use a cooler as sold.

Relative Performance and Performance per Dollar

The Relative Performance and Performance per Dollar charts have also been revamped. Specifically, the "FPU" test will not be factored in. Instead, only stock and overclocked results for "Idle" and "Blender" will be used. Maximum performance and noise normalized results have also been separated. This means users who care more about headroom and performance can see both what a cooler offers and how it does at a more acceptable noise level. The results have also been separated for AMD and Intel as both platforms are quite different, and coolers can be optimized for different platforms.

Meanwhile, Performance per Dollar is based on the maximum performance these coolers offer. As such, the stock and overclocked results where the fans are set to 100% are used. This is because we are looking at the maximum performance you can get for your hard-earned dollars. Once again, the results for Intel and AMD are kept separate for you to easily look up those for your platform of choice.