LIS2DUX12 MLC power consumption

[fibo82]

Is there any available information about excess power consumption due to the Machine-Learning Core (MLC) on the LIS2DUX12 accelerometer?

The LIS2DUX12 datasheet claims in Section 2.2 "Electrical characteristics" a current draw of up to 10.8uA (independently on ODR, which is amazing per se), but I assume this is with MLC disabled.

How does the power consumption change based on the number and complexity of the decision trees configured in the MLC?

How does it change based on the configured filters and features? Does the type of filter and features play a role? Does the Window Length play a role?

Many thanks in advance,

[Federica Bossi]

Hi @fibo82 ,

1. MLC is a computational block found in many MEMS sensors, and is expected to consume in the order of tens of milliamps, independently of the device used; however, the measure is tightly dependent on the configuration.

2. As you might have imagined, each extra decision tree configured implies more computations to perform and more power being drawn, how much depends on the size of the decision tree.

3. The real impact on power consumption, however, has to do with how many filters and features the user configures (differences among the different types is negligible), and active time of MLC. The latter depends on the MLC ODR and window length; for example, the longer the window length the less frequently we have to compute every features and update decision tree predictions.

[fibo82]

"Tens of milliamps" seems a lot: the CPU we are using (STM32L4 family) is lower-power than that. I understand the real answer depends on the "duty-cycle" of the usage of the MLC, but won't you have reference measurements of average power consumption for "standard"/basic applications for an accelerometer, such as step counting, etc.?