NVDLA inference benchmark using AGX Xavier

Hi NVidia team and anyone who could help me here:

According to the post in NVidia Developer: NVDLA Deep Learning Inference Compiler is Now Open Source | NVIDIA Technical Blog, using Xavier’s DLA core, Resnet-50’s power efficiency is 411 images/sec/W. Do any of you know how to measure the power consumption of DLA in this post? Which seems to be lower than 1 watt.

I was trying to validate this benchmark using trtexec as shown in https://developer.nvidia.com/embedded/jetson-agx-xavier-dl-inference-benchmarks. I achieved similar FPS, but not clear how to measure the power consumption of DLA as in the first post I referred.

I did follow the Jetson AGX Xavier Thermal Design Guide to measure power, but still not sure about the power of DLA device, was it for rail named “VDD_CV”? If I could measure the power consumption of DLA, was the power efficiency simple as FPS/(power consumption of DLA)?

Thank you so much!

Hi schen, yes the VDD_CV rail is what you want to measure for DLA power consumption - you have the right idea.