We upgraded our application from Jetpack 4.5 to Jetpack 5.1.2.
After that we faced performance degradation of 70msec for our cuda application.
For both jetpacks the power policy is MAXN.
In the code the component which was performance degraded based on CUSPARSE and trust pair sort of 100000K elements.
I tried my code with default Cuda 11.4 and also install Cuda 11.8 but the performance were the same.
In parallel to this there is small math code which runs in the host side
The performance combination of those is 10 msec and 150 msec of the bottle neck using jetpack 4.5.
The performance combination of those is 50msec and 220msec of the bottle neck using jetpack 5.1.2.
There is any issue about performance degradation like this?