Jetson TX2 - INA226 (Power Monitor with i2C Interface)

[Trumany]

That’s why I said “the figure 9/10 show the way of measure each power rail consumption by adding sense resistor into each branch of source rail”, VDD_CPU is output from a DC/DC which is powered by VDD_IN, the sense resistor is right in the branch line from VDD_IN to DC/DC, and VDD_SYS_CPU_IN is not VDD_CPU.

[leonardopsantos]

@Trumany

That’s why I said “the figure 9/10 show the way of measure each power rail consumption by adding sense resistor into each branch of source rail”, VDD_CPU is output from a DC/DC which is powered by VDD_IN, the sense resistor is right in the branch line from VDD_IN to DC/DC, and VDD_SYS_CPU_IN is not VDD_CPU.

OK, so if I understood you right this time, VDD_CPU supplies the CPU (all the cores?) and its source is VDD_IN through a DC-DC converter. The input of this converter is VDD_SYS_CPU_IN, which is in_voltage1_input/in_current1_input from /sys/devices/3160000.i2c/i2c-0/0-0041/iio_device/.

I’m assuming this idea is true for VDD_SYS_SOC_IN, VDD_SYS_SOC_IN and VDD_SYS_GPU_IN.

So the power comes through VDD_IN_RS, is measured, then branches out to VDD_SYS_CPU_IN, VDD_SYS_GPU_IN, VDD_SYS_SOC_IN, VDD_5V0_SYS. Is that it?

[Trumany]

Yes, VDD_IN_RS → Rs → VDD_IN → Rs → input of converter of each rail.

[leonardopsantos]

@Trumany

Yes, VDD_IN_RS → Rs → VDD_IN → Rs → input of converter of each rail.

Ah, thank you very much!

[m.samragh]

There is tons of overhead in running a bash script, and having it open, read, and close the devices for each measurement.
Just call open() on each of the devices, using O_NONBLOCK mode.
Then call read() into a sufficiently large buffer (seems like 32 bytes is plenty) each time you want to read them.

Here is a tool that reads the value a little over 4,000 times per second:

#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <unistd.h>
#include <sys/types.h>
#include <sys/fcntl.h>
#include <time.h>

int main() {
    int fd = open("/sys/devices/3160000.i2c/i2c-0/0-0040/iio_device/in_voltage0_input", O_RDONLY | O_NONBLOCK);
    if (fd < 0) {
        perror("open()");
        exit(1);
    }
    int cnt = 0;
    double sum = 0;
    double start = 0;
    while (true) {
        if (cnt == 0) {
            struct timespec tv;
            clock_gettime(CLOCK_MONOTONIC_RAW, &tv);
            start = tv.tv_sec + tv.tv_nsec * 1e-9;
        }
        char buf[31];
        lseek(fd, 0, 0);
        int n = read(fd, buf, 32);
        if (n > 0) {
            buf[n] = 0;
            char *o = NULL;
            sum += strtod(buf, &o);
            cnt += 1;
        }
        if (cnt >= 1000) {
            struct timespec tv;
            clock_gettime(CLOCK_MONOTONIC_RAW, &tv);
            double end = tv.tv_sec + tv.tv_nsec * 1e-9;
            fprintf(stderr, "Read %d values in %.3f milliseconds\n", cnt, (end - start) * 1000);
            fprintf(stderr, "average value was %.1f\n", sum / cnt);
            cnt = 0;
            sum = 0;
        }
    }
    return 0;
}

Output on my machine:

Read 1000 values in 228.846 milliseconds
average value was 15096.0
Read 1000 values in 231.470 milliseconds
average value was 15096.0
Read 1000 values in 228.912 milliseconds
average value was 15096.0

The problem is not the number of samples read in a second, but that the INA takes 2 seconds to return an idle value when the kernel ends. I think that is related to the INA3221 average mode configuration.

There is a way to change the configuration file of the INA3221 on the Jetson TX2?

Hi Mungio,

I am having the same issue with the power measurement. I can read the sysnf nodes pretty fast but their content does not seem to change much. Were you able to reduce the idle time?