I am using Xavier in MaxN mode, jetson_clocks active, Jetpack 4.3 Developer’s preview.
models are running with trtexec utility with tegrastats at background.
I found out that there is small leak at ‘CV power’ while running MobilenetV1 and MobilenetV2 standard models on GPU alone.
This ranges between 30mW to maximum of ~152mW as batch-size increases. Ideally speaking, DLAs should consume 0 mW when GPU alone is running.
This behaviour is not observed while running SSDMobilenetv1, Resnet50 and InceptionV1.
Any hint, why this would be happening?