Again. Average load on Linux is not CPU utilization.
Relax. I know. But it’s close enough. Sustained load of N where N is <= number of CPUs is roughly (but not pedantically) N * 100% CPU usage.
The permanent load of 4 thing is (or was) a reporting issue. Some 4 tasks being incorrectly reported as runnable when they were not.
@afav
How do you measure the power consumption?
I measure this with the 230V connection:
If I measure between the power supply and O6N I measure:
A =
0.5991 2.0400
0.5972 1.9350
0.5963 1.9113
V = 12.120
> A*V
ans =
7.2611 24.7248
7.2381 23.4522
7.2272 23.1650
(Measurements with plug and at the power-inlet are concurrent. Original Radxa power supply was used with a small extension to add multimeters for current and voltage. The measured voltage is mostly constant, for the current the max/cur/min was measured over a short period.)
So 25W when compiling a kernel with all cores and 7,2W at the power inlet or 27W/8W at the power plug. Currently I have 2 dongles attached (mouse/keyboard), a simple HDMI monitor and a NVME (Lexar 790NM), network: wifi installed/not working, one 8125 working at 1gbps. With extra devices (like extra SSD’s, usb devices, working wifi of ethernet at 2.5Gbps) I’d expect the maximum consumption to increase of course.
My equipment is not calibrated, but the measurements of different equipment line up reasonably well. An error margin of 10% might be reasonable, 30% seems unlikely to me.
I used an old “KILL a WATT” P4400 model, which isn’t very reliable. If I understood correctly, you got 8W in idle mode (not standby), which seems pretty good if you’re using NVMe. I think we can safely assume the Realtek 2x5Gbps card is to blame.
I’ll consider buying an O6N when they release support for the cameras (with source code).
Which kernel you use?
You mention the NVME, I have good experience with WD SN580 and Lexar NM710 and NM790. Only the WD SN520 was worse then anything I’d seen, which adds another W to the system. If you have the chance to test muliple SSD’s, do it.
While the O6N is not completly bad, the Rock5B (RK3588) operates at 4-5W with nvme connected to pcie3x4. If I wanted a real low-power system I might choose a RK3588S instead (low performance, but 2-3.5W).
As for the system, I’m currently still testing with the Radxa/Debian system (6.1 BSP). I’m planning to switch to Debian Trixie (from the other thread this week). So far I have not had much success with 6.18-kernels.
PS the O6N only has 2x RTL8125 (2.5Gbit).
RK3588S with Samsung 970 NVMe and 6.1 BSP kernel and on-chip GbE active is 1.5W (NanoPi-R6C). Without NVME, so running from eMMC is 1.1W. Measured at USB-C powerconnector (when 5V).
So depends a bit what storage speed is needed, w.r.t. CPU cores speed of S variant is the same.
If the hardware is fine, it shouldn’t be too hard to come up with a demo that proves that the O6 can be switched on while consuming less than a reasonable number of watts of power.
E.g. force the whole system into a lower power mode, force the PCI bus to a low speed, etc., just to prove the point.
Hi Rock,
just to clarify my frustration: I haven’t received my O6N board yet, so I can’t test anything myself at this point.
I got interested in the Orion lineup fairly recently (October 2025). My last ARM board was an RK3399 (Kobol NAS).
Reading through the discussions here, what strikes me is that we have plenty of observations and findings, but very rarely the full context in which they were obtained.
When two people report different results, it’s almost impossible to tell whether the difference comes from:
@hong.guo documented the cpuidle states and seems to have provided a fix for core management.
To test, you could try disabling 11/10 cores out of 12 (CPU_OFF) and see if the SoC/firmware accepts that state.
If power consumption drops significantly, it would indicate the CPUs aren’t actually reaching deep idle states.
If it stays high, it would confirm the issue is elsewhere (NoC, DDR, PCIe).
My current reading (theoretical):
ARM : Total Compute 2022 Reference Design Software Developer Guide
ACPI : Power Control System Architecture
ARM CPU Known State in PSCI Specification
Please mention the systems with the application purpose. I mentioned a system used as a desktop with mouse, keyboard and monitor (active). I tested an Odroid M2 and consumption is about 2,5W or 1,5W headless and without nvme. If I work headless it saves power, but I need the monitor showing charts 24h a day for 1-5 days (battery powered for navigation).
A Rock5B would consume 2W with emmc and 4,5W with nvme here. (Assuming both M2 and 5B fixed 12V power). Values at the power inlet without losses of the power supply.
PS you mention 5V. Lower voltage means lower consumption. When testing the Rock5B I found 0,2W lower consumption in idle and higher consumption in performance at 9V/1.7A. In general I use my systems at 12V (battery use).
The O6N draws easily 2A with 12V, at which the voltage drop is about 0,2W (on-grid, radxa 12W power supply)…
It is about idle power, so application is not relevant. Only reacting to ping maybe on the 1GbE ‘wan‘ port (or just broadcast packages), so only 2 cables connected (power and ethernet).
I can’t list application, it varies a lot, kernel compile, video play, video transcode, KVM host, etc. Also as much as possible product itself out of the box how I bought it. That is why I measure at the USB-C power input. Any loss in PSU, cables, etc should be excluded, they are there of course. Also using higher voltage via USB-C PD raises (idle) power consumption as (extra) onboard DC/DC converter is in use then and also has losses. But if your base is 12V (also mine), you will have that loss somewhere. I remember ROCK5B has 2 stages of DC/DC, not sure if in series or parallel, but look in schematics.
Potentially feeding power-hungry USB peripherals implies design choices. If no USB would be there, maybe certain stages could be removed, but then it is like 20 year old mobile phone design.
I believe that the idle power depends on the “background tasks” needed for your application. Just for the discussion:
Your statement “idle use is idle use”, my statement “your application has a minimum set of processes you need, which makes up the power consumption”.
This happens when I login over the network and stop gdm (and remove the USB-RF dongles for the mouse and keyboard while I was at it), since I don’t need X or Wayland for a NAS.
It would be even better if the gpu switched to “offline” instead of maintaining a useless blank screen, preventing the monitor to shut down. But then a blank gpu might not consume too much power.
A = 0.4990
0.4890
0.4787
V = 12.200
A*V = 6.0878
5.9658
5.8401
Behind the power supply it looks like this, with a lower limit of 6.5W, upper 7.0W and in between values of 6.9W. While the measurement might not be accurate, I think it still shows the trend I mention.
@specs You are showing a measurement that sohws circa 6.9W idle power consumption. Is this a O6N? With or without any NVMe SSD or other peripheral connected?
And then disabling gdm and removing the dongles. Monitor is still attached (black screen, no blanking).
Simply a default install of the Radxa/Debian (6.1-kernel), I’d hardly call this “optimized”. I haven’t checked the 8W with an extra digit. I think, while not precise, the results should be consistent (i.e. the measurement error should be similar).
this is well documented bug for cix p1 - with workaround that should work on radxa boards,
So far I checked on mine o6n and maybe it does not hit the bug,
I needed to compile devmem2 myself and it founds 0 events, maybe I got older bios without issue? Idle power measured to 8.5W, does not change on this fix, this looks like one of jeff g measurements for his cix board.
anybody tried that too?
