Latency calculation

[Joeri]

Hello,

I have a trained a baseline regression model as starting point for further optimization.

Before I proceed I was checking the RAM, FLASH and Latency for different boards.

When I select Nordic nRF52840 DK (Cortex-M4F 64MHz) I obtain:

When I select Arduino Nano 33 BLE Sense (Cortex-M4F 64MHz) I obtain:

If I am correct both (DK and Arduino BLE) have the same MCU, the nRF52840.

Why is there a difference in latency (163ms vs. 232ms)?

I also notice in case of nRF52840 DK there is an increase in latency (float32 vs. int8), for Arduino there is a decrease. I would expect a reduce in latency in case of quantitation, or am I wrong?
Do I overlook something?

Is it possible to provide some background about the latency calculation so I get a better understanding.

Thanks for the feedback.

Regards,
J.

[janjongboom]

Hi @Joeri, the Nano 33 BLE Sense is still using the old MACC based calculation (where we guess based on network complexity), whilst the nRF52840 uses the new benchmarks (where we run the models and then look at benchmark results ran on hardware). I’d expect the nRF52840 numbers to be closer to reality.

f32 vs i8 is interesting, will investigate.

[Joeri]

@janjongboom thanks for the feedback.

I will take Nordic nRF52840 DK as reference in further studies and perform final tests and comparison runs as soon MCU test bench is ready.

The benchmarks you refer to is there a public publication available?

(“f32 vs i8 is interesting, will investigate.” If there is a need Project ID = 77655)