The RK3399 has been working fine with the shipped fan, even without thermal paste, and I've never seen it throttle back the processors. As a result, I thought I'd trying using it with the fan unplugged, to get some idea of how well it could work with only a passive heat sink. In theory, a well designed heat-sink should work better.
I applied the shipped fan/heatsink with Arctic Silver 5 thermal paste, which I've had good results in the past, but this time removed it unplugged from the 3-pin fan connector. The computer was mounted in its supplied perspex case, and placed in a cabinet with no airflow. In order to stress the processor, I performed a Handbrake/x264 software video encode, using a 1080i MPEG-2 source. Handbrake was chosen due to its excellent multi-core implementation, which typically uses close to 600% CPU and ~50% memory, as well as the possibility to monitor the frames per second being transcoded, as a proxy for performance. I monitored two temperatures, thermal_zone0 and thermal_zone1 (referred to as zone 0 and zone 1).
As a baseline for comparison, with the fan running normally during HandBrake transcodes the temperature peaks at around 53 C (zone 0) and 43 C (zone 1), and the processors operate nominally at 1.51 GHz (the 4x A53s) and 1.99 GHz (the 2x A72s). At no stage in this test did the A53 frequencies ever drop.
+0 min: Temperatures started to climb rapidly from starting temperatures of 50 C (zone 0) and 44C (zone 1). The CPUs were operating nominally at 1.99 GHz (A72).
+3 min: The zone 0 temperature started to hit a threshold of 70 C, at which point the A72 processors started to slow, initially alternating between 1.99 and 1.80 GHz. Handbrake performance is impacted by about -6%.
+15 min: Zone 0 now ~73 C range, with processor speed continuing to step down, looking fairly stable at 1.42 GHz. Zone 1 temperature continued to climb, eventually stabilizing at around 67 C. Performance impact is about -20%.
+20 min and beyond: Zone 0 temperatures fluctuating between 72 and 74 C, A72 processors operating mostly at 1.42 GHz, and occasionally dipping to 1.2 GHz. Zone 1 temperature is remain stable at around 67-68 C. Performance hit remains about -20%.
This is sufficient to convince me that a well-designed heat sink should be able to do a pretty decent job. If anyone has any recommendations for one that would fit the board, please let me know.Author: korinel Time: 4/2/2017 01:25
I wonder if VGA coolers would be possible to mount, e.g. this minus the fan?
Author: BiatuAutMiahn Time: 4/3/2017 01:16
Why not use Firefly's fan? Or a chipset waterblock for water coolingAuthor: korinel Time: 4/3/2017 10:19
Firefly's fan is really loud.
Watercooling is fairly expensive to do well, in my experience, and it's easy to mess up and soak your computer if you do it badly.Author: korinel Time: 4/21/2017 06:31
In case this was missed by interested parties, there is now a heat sink in the store. Unfortunately it hasn't been in stock any time I've looked, but hopefully soon ...Author: isle Time: 4/24/2017 11:46
It is in stock now.Author: korinel Time: 4/25/2017 01:59
Thanks! Ordered.Author: korinel Time: 4/25/2017 02:02
I hear extremely conflicting information about how well ARM chips cope with running hot over long periods of time. Are there any temperature recommendations for to ensure that the processor isn't stressed? My processor has a fairly heavy load (all CPUs at close to 100%) for maybe 20-30 hours/week, and the rest of the time does very little other than serving files and web pages. Is there a benefit to lowering the 70 deg C temperature threshold, and, if so, how do you do it within ubuntu?Author: korinel Time: 5/7/2017 04:59
Update: I have received the heatsink from Firefly, and decided to run a relative performance test.
Methodology: I repeated the methodology above, thoroughly cleaning off the SOC with rubbing alcohol and lint-free cloth, before applying the same thermal paste. Note that the pressure on the SOC applied by both the stock cooler and the heatsink is less than on a typical PC, and so I applied some additional force in both instances to ensure that the thermal paste was distributed well. The ambient thermal environment is non-ideal (a closed cupboard), with almost zero airflow, and no precise temperature control; meteorological conditions including temperature and humidity almost certainly differed, as so there is likely some experimental error. Also, the system is operated headless, with modest background jobs, and the test involves running an intensive HandBrake software video transcoding job which utilizes almost the full capacity of each CPU core.
Idle temperature: I waited until temperatures had stabilized before starting the test, in order to ensure an accurate idle temperature. Previously, this was 50 C (CPU/zone 0) and 44 C (GPU/zone 1) with the stock cooler, operating at 408/816 MHz; In this instance it was 51/48 C with the heatsink, at the same frequencies. I do not account for the differences, which is likely within experimental error due to ambient conditions.
+0 min: Temperatures started to climb rapidly from starting temperatures of 51 C (zone 0) and 48C (zone 1). The CPUs were operating nominally at 1.99 GHz (A72) and 1.50 GHz (A53)
+1 min: The zone 0 temperature started to hit a threshold of 70 C, at which point the A72 processors started to slow, initially alternating between 1.99 and 1.60 GHz. Handbrake performance is impacted by about -6%. In the previous test it took about 3 minutes to reach this stage.
+4 min: Zone 0 now ~72 C range, with processor speed continuing to step down, looking fairly stable at 1.42 GHz. Zone 1 temperature continued to climb, eventually stabilizing at around 67 C. Performance impact is about -20%. In the previous test it took 15 minutes to reach this stage.
+6 min: Zone 1 now fluctuating between 73 and 74 C, A72 processors alternating between 1.42 GHz and occasionally dipping to 1.2 GHz. Zone 1 temperature remains stable at around 68 C. Performance hit consistently around -20%. This is about the same performance as was reached after 20 minutes in the previous test.
+10 min: No significant change, except temperature are now closest to 74/69, and the A72 processors are operating more consistently at 1.2 GHz. -24% performance hit. No major change after this.
Conclusion: The stock cooler used as a passive heat sink (unplug connector from board) offers no worse, maybe even better performance, than the purchasable heat sink. I would not recommend switching if you intend to run the RK3399 fanless; Simply unplug the connector from the fan to the board. A copper heatsink with a more complex vane structure might help, but I am aware of no available sinks that will work. Another approach might be a larger, more quiet fan providing flow across the boardAuthor: korinel Time: 6/18/2017 08:06 Last edited by korinel In 6/18/2017 08:41 Editor
Update:
I produced a copy of the top of the case using a sheet of perspex, cut with a scroll saw. In addition, I added a circular port and 4 small screw holes to that I could mount a small 40mm fan to the underside of the top perspex sheet using M3 screws, making sure that it was positioned just a little above the heat sink. The fan in question is a Scythe Mini KAZE 40mm Silent Case Fan (SY124010L), which costs about $9 at Amazon in the US. As per previous trials, I used HandBrakeCLI to transcode an MPEG-2 TS video to MPEG-4 AVC (h.264), which uses every CPU core intensively.
+0 min: Temperatures started to climb rapidly from starting temperatures of 42 C (zone 0) and 40C (zone 1). The CPUs were operating nominally at 1.99 GHz (A72) and 1.50 GHz (A53)
+3 min: Temperatures have climbed to 64 C (zone 0) and 54 C (zone 1), and CPUs continue to operate at nominal temperatures.
+10 min: Zone 0 now peak at ~68 C, with processors still functioning at maximum frequency. Zone 1 temperature continued to climb, eventually stabilizing at around 57 C. No performance impact.
+15 min: Zone 0 now peaks around 68-69 C, and zone 1 at 57-58 C. Still no processor throttling or performance impact. After this point, no further temperature increases were experienced.
Conclusion: Positioning a fan above the heat sink, even a small, quiet one, makes all the difference. For those that are happy cutting their own materials, I recommend it. Positioning the fan as close as possible to the heat sink makes a large difference. Just an extra 15 mm above where I settled on reduced the effectiveness to almost zero.
Additional: The fan is actually rated at up to 12 V, but I was using it at 5 V due to the even-more-silent operation. However, I discovered that at 5 V it often struggles to spin up after a power cut, however, sometimes requiring a nudge. Switching it to 12 V made it a little louder, but still very quiet, and the cooling performance improved massively, peaking around 54 C and 45 C. Also, after conducting the initial test I lowered the fan even closer to the heat sink, which further improved cooling performance. The 12 V to 5 V trick to get 7 V might also be worth trying out for an intermediate noise level. In any case, this fan is pretty awesome for its size, and WAY quieter than the stock cooler which whines unpleasantly.
