1- Jetson nano doesn’t support INT8, why we use INT8 for jetson nano in the FaceDetection-IR? In my opinion, the conversion FP32 into INT8 should be achieve like FP16 on jetson nano, right?
2- Because this device doesn’t support INT8, Is it possible to run INT8 models on this device? If so, Is this precision mode has effective on jetson nano with improving in latency?
1.Yes, Jetson does not support INT8. For the figure inside FaceDetection-IR https://ngc.nvidia.com/catalog/models/nvidia:tlt_facedetectir, it already mentions “On the Jetson Nano FP16 inference is run”. Please read it again.
2. It is not possible to run Int8 precision on Nano.
Thanks so much.
So the FaceDetection-IR can be achieved 103 FPS on jetson nano in FP16?
Yes, according to that link.