Why the large performance discrepancy between DS 6.2 and DS 6.0

[steven.white1]

Please provide complete information as applicable to your setup.

**• Jetson Xavier AGX
**• DeepStream 6.2
**• Jetpack 5.0.2
**• TensorRT 8.4
• Issue Type( questions)

Looking at the DeepStream documentation, the performance numbers between DS 6.0 and DS 6.1/6.2 drop greatly. 195 fps in DS6.1/2 down from 411 fps in DS 6.0 for the same Tao pre-trained PeopleNet model for example on the Jetson Xavier AGX.

https://docs.nvidia.com/metropolis/deepstream/6.0/dev-guide/text/DS_Performance.html#tao-pre-trained-models

https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_Performance.html

Are these results correct or are there differences in the implementations that would explain the performance drop off?

[mchi]

Hi @steven.white1
Thank you so much for reporting this!

Will check and get back to you this week.

[mchi]

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

Hi @steven.white1 ,
It’s because the model itself has changed - heavy and more accurate. Back then DS 6.0 we used peoplenet 2.1 and now we use peoplenet 2.6 which is heavy - so it is expected.