Lower perfomance when trying to replicate the DLA Dense Performance results

I have a Jetson AGX Orin 64GB and I’m testing it concretely on the Orin Dense Performance section of the page. I have downloaded your models and used your commands lines provided on the README.md in /scripts/prepare_models/.

Logs on verbose mode are here:

log_retinanet_resnext50_MAXN.txt
log_retinanet_resnet34_MAXN.txt
log_ssd_resnet34_MAXN.txt
log_resnet50_MAXN.txt
log_ssd_mobilenetv1_MAXN.txt

• RetinaNet ResNeXt-50: yours is 78 fps, mine is 39 fps
• RetinaNet ResNet-34: yours is 108 fps, mine is 53 fps
• SSD-ResNet-34: yours is 83 fps, mine is 41 fps
• ResNet-50: yours is 2037 fps, mine is 504 qps * 2 (batch) = 1008 fps.
• SSD-MobileNetV1: yours is 2664 fps, mine is 655 qps * 2 (batch) = 1310 fps.

My results seems to be constantly around half of your reported results. I copy and paste your command lines for execution so I don’t think I’m missing an option here. I double checked that I was on MAXN power mode. I do not understand what I’m missing.

Thanks in advance!

Start with the Orin Dense Performance section of README.md and the commands in /scripts/prepare_models/. Compare the reported commands and attached verbose logs for the five models against the documented benchmark setup. Done means identifying the configuration or option causing the roughly 2x gap, or explaining why the results cannot be reproduced.