[FAQ] TDA4VM: How to perform power benchmarking of model inference/run on TDA4VM SoC

Hi,

Prerequisites :

• EVM Board
• SD card with processor 8.6 SDK RTOS is created as per:
  https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/vision_apps/docs/user_guide/BUILD_AND_RUN.html

The power measurements are done using the INA226 Linux driver. Using the sysfs interface we can measure the power consumption on all the 32( 2 sets of 16 rails) voltage rails. We can measure 16 rails (1 set) at a time using the I2C mux.

One can capture the snapshot using the below steps on the target EVM follow below mentioned steps.

Read file /sys/class/hwmon/hwmon8/power1_input to get the power reading as shown below, the readings are in micro watts unit.

root@j7-evm:~# cat /sys/class/hwmon/hwmon8/power1_input

You can run/inference any supported model then read the above mentioned file to get power reading.

Example : Inferencing Object Detection model from vision apps directory

root@j7-evm:/opt/vision_apps# ./run_app_tidl_od.sh

Read the /sys/class/hwmon/hwmon8/power1_input file before doing model inference and note the power reading, lets say its X micro watts, then run the model inference as mentioned above and take reading lets say its Y micro watts while model inference is running, the delta (Y - X) is your final reading.

Once can do power benchmarking for Object Classification or Segmentation by following above steps, by invoking inference run to these models.

Regards,

Pratik