Orin NX performances - comparing with Xavier NX

[Pelepicier]

Hi everyone,

I worked before with the Xavier NX (8gb) and I bought last week a Orin NX (recomputer from Seeed), 16gb version.
Testing the performances of the board I have same results than on a Xavier NX (sometimes better with the Xavier…), with SSD Mobilenet (INT8 and FP16) and yolov5 (s and m6 for now). I’m very disappointed about these results, has anyone made a comparison of the 2 boards?
I checked the GPU activity which works at around 90% all the time.

Also when using Jtop I can see the 2 DLAs in off mode, how can I enable them?

Thanks

[AastaLLL]

Hi,

Based on our benchmark results, OrinNX can reach 2517.99 samples/s on ResNet-50 and Xavier is 1245.10.

Jetson Benchmarks

Do you test it with the same power mode or MAXN?
For example:

$ sudo nvpmodel -m 0
$ sudo jetson_clocks

Thanks.

[Pelepicier]

Hi @AastaLLL ,

Yes it works better now, I was in the mode 25w shame on me ;)
I have approximately twice better performances now, can you tell me how to activate the 2 DLAs ?

[Pelepicier]

Also using the MAXN mode I have this message : “system throttled due to overcurrent”. Is there any risk for the board ?

[AastaLLL]

Hi,

Do you run the inference with TensorRT?
If yes, you can put the model to DLA by using --useDLACore=[ID].

When the throttling message occurs, the system will reduce the clock automatically to protect the device.
If most of your tasks run on GPU, you can set up a manual clock mode and adjust the CPU clock to avoid the message.

Thanks.

[Pelepicier]

@AastaLLL thanks I could generate the engine file for DLA with trt, I’ll try it soon.