What and Why

Its already well-known that most computer component manufacturers do a poor job of applying thermal compound, often using a material with low thermal conductivity, and then applying too much of it. In the pursuit of better cooling and higher performance, a large number of enthusiasts will replace the thermal compound between their CPU (or GPU) and heat sink with a slightly better thermal grease, and also (try to) do a better job of applying the right amount. I have done this a handful times with different types of hardware, and have seen a range of non-existent to pretty good results.

In the last couple of years, ‘liquid metal’ thermal compounds have been coming into the mainstream. They promised both better heat transfer and greater risk to hardware. These are generally products based on gallium, a metal which is liquid at a little above room temperature (and won’t kill you like mercury). I decided to try one of these liquid-metal products out on my laptop and report the results.

Evidence of Thermal Throttling

Before making any changes, I ran some benchmarks and took some measurments. The most interesting one, to me, is the plot of my laptop’s GPU temperature overlaid on a plot of the GPU Clock speed during a 10-minute load test.

Here, you can see that the GPU clock speed increases to just over 1 GHz in response to the sudden increase in load. At the same time, the GPU Temperature begins to increase as well. As soon as the temperature hits some limit, in this case, 87 °C, it lowers its clock speed to avoid overheating (and causing damage). Throughout the test, it appears to lower the clock speed several times to keep the temperature below/at some limit. When the test ends, there is no more load on the GPU, and the clock speed it adjusted lower to save power, and the temperature drops back down over a couple of minutes.

I think this chart does a good job of showing the balancing act that the GPU’s frequency scaling system has to do to maximize performance, minimize power consumption, and maintain a safe operating temperature. In theory, if we can improve cooling in some other way, then the GPU will not have to sacrifice its clock speed (and performance) to reduce the temperature.

Replacing the Stock Paste with Liquid Metal

The product I used was Thermal Grizzly Conductonaut, which is a mixture of tin, gallium, and indium. The manufacturer claims that it offers “a very high thermal conductivity”. According to the wikipedia article on thermal grease, the range of thermal conductivity for regular thermal grease is 2-8 W/mK. Thermal Grizzly states that their product has a thermal conductivity of 73 W/mK. For comparison, copper, which most heatsinks are made out of, has a thermal conductivity of about 400 W/mK.

After disassembling the laptop and removing the heatsink, we can see that the CPU (right) & GPU (left), have way too much factory applied thermal paste, as evidenced by how much of it has spilled off the die.

I cleaned it all off with alcohol and coffee filters.

Since liquid metal is conductive, it would be bad if any of it spilled on the components around the die, so I applied a couple of coats of conformal coating on to the area surrounding each chip as insulation.

Then I squeezed a small dab of the liquid metal onto the CPU & GPU, and spread on as a thin layer using a q-tip. I also put a thin layer on the heatsink (not pictured) that would be in contact with the chips.


With the compound now applied, I put the heatsink back on, screwed everything back in place, and was pleased find that my expensive laptop still booted.

Testing Tools/Procedure

To test the performance effects of this modification, I used the Unigine Valley Benchmark. It is familiar to most people who build PCs and do benchmarking. The load is GPU-heavy, CPU usage sat at around 5 - 15%. I ran the benchmark twice for each configuration, so I could measure the performance drop from the GPU throttling due to the GPU heating up during the first run.

A recording of the valley benchmark
Metrics, such as GPU temperature and clock speed were recorded using [MSI Afterburner](https://www.msi.com/page/afterburner). This is another popular tool that allows users to change GPU settings such as clock speed and voltage. I did not make any settings changes, I only used the ‘Log history to file’ feature in the metrics window, which saves timeseries metrics to a CSV.
png
MSI Afterburner
I used the following procedure before and after replacing the thermal compound: 1. Start recording stats 2. Let the laptop idle for 1 minute 3. Start Unigine Valley and run the benchmark twice 4. Stop Unigine Valley and idle for 2 minutes 5. Stop recording stats

Results

The results were good. After the modification, temperatures were lower both during idle and load, and there was no thermal throttling of the GPU.

Temperature & Clock Speed



The improvement in GPU temperature here is large both during idle and load. On average, the temperature is 13.1 °C lower after applying the liquid metal. On the second chart, we can see that there is no thermal throttling whatsoever after replacing the thermal compound, and the GPU clock speed stayed at its max of 1034 MHz the during the entire benchmark, unlike before when it throttled repeatedly to keep the temperature down.

Although the benchmark caused little CPU load, there was an improvement in temperature there as well. I suspect this has more to do with the fact that the CPU & GPU in my laptop share the same heatsink/heatpipes than with heat generated by the CPU itself. Unfortunatley, I did not run a CPU load test for the before & after comparison.

Performance Benchmark Results

As I mentioned earlier, I always ran the benchmark sequence twice in a row, to measure the rendering performance in terms of min/max/avg frames per second over time when under load.

In the “before” runs, performance starts out good, but takes a very noticable hit by the time the second run occurred. We can attribute this to the GPU heating up during the first run and throttling. When the second run starts, the temperature is already high and the clock speed is already reduced.

The FPS in the “after” run was consistently higher, and rendering did not suffer on consecutive runs.

Conclusion

In conclusion, I would say that switching to a liquid metal thermal compound was well worth the trouble. It solved a real thermal throttling issue, increased performance, and was easy to do. The cost was low, under $30 of materials (thermal compound, conformal coating, alchohol), and I estimate that I have enough materials left over for 3-5 more applications, that I will likely never find a use for.