Hi,
I'm encountering an issue during the optimization process using Edge Cloud AI. Sometimes, even when using the same model architecture but with different file sizes, the optimization results vary significantly. In some cases, especially with smaller models, the NPU on the STM32MP257-DK is barely utilized (as shown in the benchmark below). Paradoxically, I’ve observed worse optimization performance with lighter models.
In the images below, I show an example: from left to right, the models progressively increase in resolution, yet the optimization results behave inconsistently. Any suggestions on how to address this?
TINY_YOLO_V2_192_INT8.TFLITE:
Inference time: 87.52 ms
GPU: 95.85%
NPU: 4.15%
TINY_YOLO_V2_224_INT8.TFLITE:
Inference time: 5.920 ms
GPU: 3.54%
NPU: 96.46%
TINY_YOLO_V2_416_INT8.TFLITE:
Inference time: 12.13 ms
GPU: 4.50%
NPU: 95.50%
