Our test platform uses an Intel i9-7960X 16-core CPU, one that we reviewed overall positively last year. The CPU is clocked to 4.6GHz at 1.22V and has all over-current protections and other power limitations disabled. The motherboard is a Gigabyte X299 Gaming 9. The 7960X has been delidded and is using Thermal Grizzly Conductonaut as its liquid metal, with the leftover silicone adhesive still in place on the double substrate. We used Thermal Grizzly Kryonaut for the thermal paste between the IHS and our cooler, a Corsair H115i with two Noctua NF-A14 fans. We also used GSkill Trident Z Black memory at 3600MHz across 4 channels. The power supply is a Corsair AX1600i in single-rail mode, as we’re pushing about 40A during our test.
For testing, the system ran FFT workloads with automated cycling for the past year. Rather than thermally strain the CPU 24/7, we alternated loads with break periods, more similar to a real use case, which will also accelerate aging with rapid expansion and contraction of the liquid metal and thermal interfaces. Simulating age in electronics is best done by thermal cycling, not by sustained loads. Final thermal testing for results measurement uses a GN Blender benchmark that was purpose-built for CPU load, but this is not what was used for the year-long thermal torture.
We ran a few thermal numbers that are most critical:
- Initial testing before the aging process
- Testing after and with no changes
- Multiple tests after, but with the thermal paste replaced
- Multiple tests after, but with new liquid metal
Limitations of Testing
There are some limitations of testing with this. For one of them, we’re only using one sample, so this isn’t globally conclusive of all liquid metals, or even of Conductonaut. This gives us a great baseline and a case study, but we would need more samples to really make hard claims. The liquid metal isn’t the only thing that aged – there’s also aging of the thermal paste and, although this one is unlikely to impact in any meaningful way, the cooler. To partially account for this, we did some tests with new thermal paste applied both before and after replacing the liquid metal.
Results – How Often Should You Change Liquid Metal?
For results, we measured the original application at 62.9 degrees Celsius delta T over ambient. That’s data from one year ago. Today, without changing anything, we measured thermal performance as 63.3 degrees Celsius over ambient. This is well within our normal +/-1C error margin for testing and, given that we have a whole lot more variables with a one-year gap between tests, including potential for cooler, paste, motherboard aging, and higher chance of technician error, these numbers are close enough that we cannot state any difference. There was no significant change. The change we are observing is within error and test margins, and if not error, it’s more likely resultant of paste aging.
So for that, we applied new thermal paste, but kept the CPU socketed and didn’t touch the liquid metal application. With new paste applied and the old liquid metal, we measured performance as just under 65 degrees over ambient. Again, this is within reasonable error. We are also using a different tube of Kryonaut this year versus last year, so that could factor into the results.
With the new liquid metal application, we measured performance at around 63.5 degrees, which is within reasonable error of both our original result and our new thermal paste result.
Conclusions on Liquid Metal Aging
We had other liquid metal application results that didn’t make it to the charts: The tests above are from our third of three attempts at re-application. It can sometimes take a few tries to really get a good application of liquid metal, primarily due to surface tension and quantity of liquid metal used. We wanted the most level and stable per-core temperatures, and so took the best of each application.
The point of detailing this is that, of three new applications, we had two results that we threw away. When conducting the original test, we did the same thing: We had three applications, and then the third was the best one. This shows that there’s more variance in application of liquid metal – a whole lot more – than there is measurable difference in liquid metal aging on our platform (again, based on a case study – please note that this is not 100% conclusive for all LMs on all platforms). You should be more concerned about how well the liquid metal is applied than about its aging. That said, monitor thermals every now and then just to make sure all is well. If you’ve been running liquid metal in a system for a while, please let us know how it has aged.
There very likely is some sort of aging, but we were not able to measure it in this testing. It could be at the two-year mark, it could be different liquid metals, it could be the testing, but we did not see any meaningful change after one year of performance testing. There is far greater impact from changing the method of application, which will happen every time LM is re-applied. Most likely, if re-applying liquid metal, simple experience of application will cause more improvement than the change itself.
Editorial, Testing: Steve Burke
Video: Andrew Coleman