After installing the Enzotech CNB-R1 Northbridge heat sink, I ran a temperature test using the CUDA Smoke Demo. Basically start up the demo, and take temperature readings over the next 30 minutes or so at 5 minute intervals. At the end of the demo close the Smoke window. Afterwards I measured the temperatures after 1 minute, 5 minutes and then a final reading after 10 minutes. This tells me how quickly the CPU and the GPU heat up, and how quickly they cool down after running a graphically intensive program. I’m sure there’s a more professional and standard way to do this, but I like seeing smoke when doing temperature testing. I think it’s more fun that way.
Note that this is just a good stretch of the legs for the Jetson, the smoke demo uses two ARM cores at around 75%. The Jetson actually has 4 cores (5 if you count the “shadow” core) so the CPU could actually be cranked higher still and generate more heat. Still, lots of activity going with the GPU doing a little dancing on its own.
The Arctic Silver Ceramique thermal paste has a break in period of approximately 25 hours. Note that the break in period is cycled, the expectation is that there are periods of high activity and low activity. Note that it’s not adequate just to turn it on for 25 hours and expect it to be broken in. I tested right after installation, about 8 hours into the burn cycle, and the latest at 20 hours. Note that things could still have a chance of getting better as the compound will thicken over the next 100 to 300 hours. Tests are all run with the ambient temperature being 26.7°C. At least that’s what the thermostat on the wall says, summer here in California and all. Anyway, here are the results:
Here’s the executive summary: With the stock heat sink and fan things top out at around 50 °C, with the Enzotech CNB-R1 it rides up to around 91°C. Note: I’ve seen notes that critical temps are around 101 °C for the CPU and GPU. Over the burn in period, the temperature has not changed in any significant way. I did get brave and let it keep running for more than 20 minutes to find out where the temperature plateaus are with the Enzotech in the second and third runs.
Assuming that the thermal trips work (always a tricky question on early production development boards), it’s probably ok to run it with the new copper crown, but I can’t say I’d be comfortable starting a really compute intensive task and walking away for an extended period.
I’ll inject here that my own personal internal temperature is around 37 °C, and doesn’t fluctuate that much even when I think really hard. Just sayin’.
About the break in period and all that? I haven’t seen much change. It’s entirely possible that the thermal paste wasn’t properly applied, but we’ll save that for the post mortem. At this juncture I wouldn’t recommend this upgrade if you’re planning to really crank out the GFLOPS.