Power consumption and SoC temperature in some edge conditions

[egil]

We attached the following peripherals to a M1 board and logged power consumption and CPU temperature.
- eMMC Module
- M.2 NVMe SSD
- SATA SSD
- MIPI CSI Camera
- MIPI DSI Display Vu8M (with maximum backlight brightness)
- 4K HDMI TV
- WiFi USB dongle (WiFi Module 5B)
- USB mouse and keyboard

Software setup
- Android 11 on eMMC
- CPU and GPU stress test app showing on the Vu8M
- MIPI CSI Camera preview app showing on the HDMI TV
- Continuous heavy 'dd' access from/to NVMe and SATA SSD devices in background
- WiFi 5GHz 11AC connected "wireless iperf" loop test

Power source : SmartPower3 for power transient logging

We ran the relatively heavy load test for a few hours.
The average power consumption was 13.5Watt and the CPU temperature was 61°C. Ambient temperature was around 22~24°C.
Observed peak values were 15.7Watt and 63°C.
Therefore, we believe the recommended 12V/2A (24Watt) PSU and the stock heatsink must be sufficient for typical applications.
When we didn't connect any peripherals and ran a CPU stress program on a headless configuration, the power consumption was 4.5Watt and the SoC temperature was 51°C.

M1 Power Comsumption Graph.png (60.48 KiB) Viewed 328 times

[egil]

Have you tried measuring headless idle consumption with cpu frequency scaling? I've configured cpufrequtils to scale down to the lowest 408mhz, but I don't know if it has a measurable impact or only diminishing returns.
My plan is to use it as a streaming host with Jellyfin primarily for the kids, so it will be idling most of the time.

[egil]

Have you tried measuring headless idle consumption with cpu frequency scaling? I've configured cpufrequtils to scale down to the lowest 408mhz, but I don't know if it has a measurable impact or only diminishing returns.
My plan is to use it as a streaming host with Jellyfin primarily for the kids, so it will be idling most of the time.

The difference might be very negligible since Cortex-A55 on a 22nm wafer is quite power efficient.
It was around 1.1~1.3Watt if governor was ondemand or interactive. There was no USB device connected.

[mctom]

Did you update SmartPower3 firmware prior to the test?

How was the board oriented during the test? Personally I want to check the impact of upright positioning, I think this may help significantly, especially in passive cooling scenario.

Have you tried measuring headless idle consumption with cpu frequency scaling? I've configured cpufrequtils to scale down to the lowest 408mhz, but I don't know if it has a measurable impact or only diminishing returns.
My plan is to use it as a streaming host with Jellyfin primarily for the kids, so it will be idling most of the time.

The difference might be very negligible since Cortex-A55 on a 22nm wafer is quite power efficient.
It was around 1.1~1.3Watt if governor was ondemand or interactive. There was no USB device connected.

I switched to interactive and ondemand, but there was no visible difference in power consumption, or CPU frequency, even if M1 was idle. Probably governor parameters need some attention. I'll look into that, as I'm interested in battery operated application.

[egil]

I switched to interactive and ondemand, but there was no visible difference in power consumption, or CPU frequency, even if M1 was idle. Probably governor parameters need some attention. I'll look into that, as I'm interested in battery operated application.

By default both min and max = 1.99Ghz. You need to set minimum to a lower state to get any effect. eg

Code: Select all

cpufreq-set --min 408

to enable the lowest state.

[mctom]

Did you update SmartPower3 firmware prior to the test?

We used the latest V1.9 (Feb-22nd) version.

[mctom]

I ran a quick test (ok, 2 hours..) to see if orientation matters.

2022-04-08-143255_1915x856_scrot.png (106.87 KiB) Viewed 239 times

Unit had only Ethernet plugged in, has NVMe drive.
I ran stress -m 4 and kept this load constant throughout the test. I registered temperature in thermal_zone0 in 1s intervals. Red plot is smoothed data with single pole low pass.
Yellow graph serves to show 5 sections of the measurement:
1. Unit lies on a heatsink, on a naked desk
2. Unit stays upwards
3. Unit lies again as in #1
4. I put 10mm isolating foam under the unit
5. I positioned unit vertically again, on a foam.

It wasn't apparent during the experiment, but the periods were too short to stabilize temperature. But trends are still clearly visible:
1 -> 2, temperature drops
4 -> 5, temperature also drops, by at least 3 deg C!
How's that a big deal? Well, cooling (in)efficiency is often described as thermal resistance from chip core to ambient air, which is calculated as difference between two temperatures, divided by output power.
If I just showed one could decrease core temperature from 50 to 47 degrees by merely rotating M1, that means the reduction of thermal reisistance by...
1 - [(47 - Ta) / (50 - Ta) ], where Ta I assume generously it's 24C... 12%.
I measured some 5.3W of power draw during my test. It is safe to assume that in case of 11W the temperature difference between two orientations will be double, some 6-7 C.
I know it's oversimplification, that assumes all heat dissipation comes from a CPU, but upright position promotes cooling of almost everything on a board likewise.

Moreover:
3 -> 4, temperature starts to rise faster.
That means a table top did play a role in absorbing heat. Adding isolating foam made the temperature climb faster, which means the target temperature was higher (because thermal capacity was the same).
That means M1 temperature measurements are affected by the surface it has contact with - as with everything else.

Just to make matters clear, I think M1's heatsink is more than enough for any possible application. That's just my 3 grosze to do a tiny extra to prolong its life and reduce power consumption.