News Enthusiast Develops First Low-Profile Raspberry Pi 4 Cooler With Direct-Contact Heatpipes

[CooliPi]

Cool, thanks for the info about stress-ng ! I'm trying it now. Regarding glmark - it stopped compile with an error. No time to investigate now.

I also know dd and netcat, good point.

So far, with stress-ng running, the BCM temperature is 50degC, ambient 28degC. Frequency 1850MHz, over_voltage=2, one FullHD HDMI output.

I'll compare it with linpack, to test which one is a bigger power hog.

Linpack load increases the BCM temperature 10degC more than stress-ng

Maybe some stress-ng subtest has a bigger power consumption impact than the one I've chosen.

 

[bit_user]

I also know dd and netcat, good point.

The main thing about dd is to read from the raw device. That will really stress the USB3 controller, since it won't be bottlenecked by the kernel's filesystem driver.

 

[CooliPi]

Good point! I'd almost forgot filesystems are cached. dd_rescue to the rescue. Its verbose mode allows to see throughput in real time.

I've tried to update my system and overclock, but overclocking the gpu with over_voltage=3 and >=600MHz leads to stuck graphics and also the system right after displaying the X screen. Maybe too low over_voltage. I fear instability with higher voltages and frequencies.

I've run linpack with over_voltage=3, arm_freq=1850MHz, gpu_freq=600MHz and after some time, when BCM reported 60degC it'd reset. With passive cooling. So this is a sign of the PMIC overheating. I'll try with that huge 60mm fan.

I need to add a thermal bridge for the components around the PMIC, these are inductors that hold energy. When they overheat, their ability to pump energy decrease. Only after that can I tell whether it's the PMIC itself or the inductors. PMIC performance is definitely temperature dependent.

So for a good overclock, low PMIC temperatures are essential.

 

[CooliPi]

I've tried to have 4 threads of linpack running and run some WebGL demo running in Firefox. It almost seems like loading the GPU takes away cycles from Linpack, leading to marginally lower temperatures - about 1degC.

It'd be best to let the GL test render into some huge framebuffer, say 4k, to load CPUs the least it could. But it depends on its command processor. If it renders without a CPU assistance, then it'd make sense.

One last test I could do would be with a camera, recording in MPEG4. It runs almost entirely in the GPU (VPU). But some mpeg packet assembly could take away cycles from the main power hog - linpack - too.

I've yet to test it.

 

[ColinIanKing]

The matrix stressor is generally a good stressor to try because it exercises CPU/cache/memory.

stress-ng --matrix 0 --matrix-size 256

..adjust matrix-size to different sizes to see how that affects the temperature. You can also swap the matrix ordering to be less cache efficient and this may help make the core hotter too. Use the
--matrix-yx option for that, e.g.

stress-ng --matrix 0 --matrix-size 256 --matrix-yx

The --matrix 0 option will run the matrix stressor on each of the on-line CPUs.

Some devices have thermal zones, so the --tz option may also be useful for getting thermal temperature information.

 

[bit_user]

You can also swap the matrix ordering to be less cache efficient and this may help make the core hotter too. Use the
--matrix-yx option for that, e.g.

stress-ng --matrix 0 --matrix-size 256 --matrix-yx

Swapping to column-major access (for large, especially power-of-2 matrices) is going to stress the core less, because it'll be stalled on memory accesses.

Probably a good way to stress memory, though. But, that's a different test case.

 

[CooliPi]

Seems like my stock (part of a distribution) stress-ng is perhaps not compiled with NEON instructions, because it can load it to max. 48.3 degC. With linpack, I can get to 60degC.

Ambient temperature is 28degC

This stress-ng temp was for the largest array (8192x8192)