Part of Code executing on CPU of Jetson Agx Orion board is taking more time compared to execution time of same code on Jetson Agx xavier board

Accelerated Computing CUDA CUDA Programming and Performance
1

I developed a code in which some part is being executed on CPU and some part on GPU. Initially I tested it on 32GB Jetson Agx Xavier Board on which execution time of the part running on CPU was around 150 microseconds. When I tested the same code on 64GB Jetson Agx Orion board, the part executing on CPU is taking around 500 microseconds. I want to know why this is happening while Orion board has better clock speed. Is there any solution to this as i need execution time to be less.

I have already executed command
sudo nvpmodel -m 0

The part of code executing on GPU is giving great performance on Orion board

2

I suggest asking this question on the relevant Jetson forum.

3

Could be something as 8 MB L2 vs. 3 MB L2. But as you wrote, not GPU related.

4

Possible but seems unlikely? Other than for this one metric, the “speeds & feeds” for both CPUs suggest that Orin should be faster than Xavier.

The size of the performance difference suggests a debug build versus a release build. Other than checking build settings, profiling should reveal the root cause of the performance difference.

5

Thanks for replying.
I am executing code on both the system with same simple command as-
nvcc Filename.cu -oTestOutput.

I will try to do profiling

6

You could try providing optimization flags directly to the host compiler with -Xcompiler. For example, for Orin it might be -Xcompiler "-O3 -march=armv8.2-a -mtune=cortex-a78ae"

Please note that experience of participants in this sub-forum with Jetson platforms tends to be limited, and you will likely get much better and more relevant answers by asking in the Jetson forum already linked by @Robert_Crovella above.

7

Thanks!!
I will try to set optimization flags.
I have Posted this issue on Jetson forum.