This thesis evaluates the performance of the XIAO ESP32-S3 Sense development platform for edge artificial intelligence and TinyML applications. We
consider keyword spotting, image classification, and object detection, and
compare FP32 and INT8 model variants for each application. The impact of
quantization is evaluated using inference latency, energy consumption, and
estimated RAM requirements. The results show that INT8 quantization has
the greatest effect on the image-based applications. For image classification,
latency is reduced by 96.18%, energy per inference by 96.10%, and RAM
requirements by 62.55%. For object detection, the corresponding reductions
are 95.73%, 95.85%, and 71.80%. For keyword spotting, the reductions in
latency and energy are smaller, while the overall RAM requirement remains
unchanged. The results show that INT8 quantization on the ESP32-S3 can
substantially reduce computational and energy requirements, while the magnitude of the improvement depends on the characteristics of the individual
application.
|
