Pi3 incorrect results under load (possibly heat related)

[jahboater]

Most people do not want to use a fan.

My difficulty, and those of some others, seemed to be related to transients caused by suddenly switching from one core active to all four cores. This is why over_voltage was needed even when running at default frequencies. Of course throttling is no fun either, but much less of an issue than incorrect results and system crashes.

Can the ondemand governor de-activate cores? That's the job of the hotplug governor:

Hotplug Governor:
The Hotplug governor performs very similarly to the OnDemand governor, with the added benefit of being more precise about how it steps down through the kernel's frequency table as the governor measures the user's CPU load. However, the Hotplug governor's defining feature is its ability to turn unused CPU cores off during periods of low CPU utilization. This is known as "hotplugging."

[ejolson]

Can the ondemand governor de-activate cores?

You are right, I didn't technically mean activate and deactivate. I've edited my previous post for clarity to use the term busy instead of active.

In the linpack test all four cores are active, one is busy, the CPU frequency is currently at 1200 and the voltage is already stepped up. Then the other 3 core suddenly become busy and the system crashes unless I've set over_voltage=2. It is possible the rendering engine in the Chromium browser supplied with Pixel Raspian is multi-threaded in such a way that a similar transition from one to four cores busy also regularly happens.

[alexdev]

After increasing over_voltage to 2 my Pi3 reached 6.5 GFlops with stock frequency and N = 8000.

With N = 10000 I get 6.770 GFlops. That's a bit strange. At the beginning I got around 5.3 GFlops, but after repeating the benchmark several times the perfomance improved.

Overclocking the Pi at 1300 GHz results in a failing benchmark, no matter how much I increase the over_voltage value. The GFlops rate is also low (around 5.2 Gflops).

I have a small 5v FAN attached to the 5V rail of my Pi, and I observed that while running the cpuburn-a53 stress test the FAN speed decreased and after stopping the stress test the FAN speed reached again full speed. So maybe the fan is draining too much current from the PSU and this is why I need to change the over_voltage value.
I think it's better to power the fan externally.

[ejolson]

I have a small 5v FAN attached to the 5V rail of my Pi, and I observed that while running the cpuburn-a53 stress test the FAN speed decreased and after stopping the stress test the FAN speed reached again full speed.

It seems a bad sign that 5V power drops enough for the speed of the fan to change.

I've decided to run my Pi B3 underclocked to 900MHz, because reliability is more important for the intended application than performance. If you underclock your Pi, does the fan still slow when you run the stress test?

[alexdev]

I have a small 5v FAN attached to the 5V rail of my Pi, and I observed that while running the cpuburn-a53 stress test the FAN speed decreased and after stopping the stress test the FAN speed reached again full speed.

It seems a bad sign that 5V power drops enough for the speed of the fan to change.

I've decided to run my Pi B3 underclocked to 900MHz, because reliability is more important for the intended application than performance. If you underclock your Pi, does the fan still slow when you run the stress test?

Yes, I can hear that the fan is not spinning at the same speed everytime I start/stop the stress test.

I have underclocked it to 900MHz but the problem persists.

The fan is also making some weird noises sometimes (i have to rotate the board a bit to stop it, cheap chinese thing) but I don't have another solution until summer when I'll build a custom wooden case for my project and I will also build a proper cooling system for the CPU.
There is no blinking red light on the Pi or any lightning bolt on the screen, so I think it's OK.

[ejolson]

Yes, I can hear that the fan is not spinning at the same speed everytime I start/stop the stress test.

An audible indication of how much computation is being done can be amusing. Vintage computers create enough RF interference that you can tell what the computer is doing by tuning to a unused spot in the AM band. I've always wanted an animated stream-punk wallpaper with stream engine noises that changed based on load level, CPU utilisation, network traffic and SD card activity. If you included a few brass fittings with your wooden case it might just work. However, someone still has to design the wallpaper and write the sound synthesiser for the steam engine. I wonder if Sonic Pi could be used for the synthesiser.