Jetson Orin AI Performance

[kjasper]

Hi, is there a comparison of the TX2NX and Jetson Orin Nano in the same units? Why don’t the numbers for Orin use TFLOPs?

[AastaLLL]

Hi,

We don’t have the comparison between TX2 and Orin but some data that compares Xavier and Orin.

Jetson Benchmarks

Orin has both Tensor Core and GPU Core. Tensor Core is measured in TOPS.

Thanks.

[dusty_nv]

It’s because TX1/TX2/Nano support FP16/FP32 hardware (TFLOPS), whereas Xavier/Orin support INT8 hardware (TOPS) in addition to FP16/FP32.

[kjasper]

Are there numbers for the FP16 performance? If I run a FP16 model it won’t run on tensor cores so its a direct comparison correct? How about video encode performance? It just says it can encode on a CPU but not what resolution or codecs.

[dusty_nv]

I don’t believe that we have published FP16 benchmarks for Orin Nano. The video encode performance depends on which library and settings you are using (for example, the x264enc GStreamer element). What I would recommend for the time being, is to use Jetson AGX Orin to emulate the Jetson Orin Nano configuration to run the benchmarks that you need:

Develop for All Six NVIDIA Jetson Orin Modules with the Power of One...

[kjasper]

OK, does the Jetson Orin Nano 4GB support 20-bit bayer camera input?

[dusty_nv]

The MIPI CSI Camera Interface section of the Orin Nano datasheet lists the following formats:

• Supported input data formats:
o RGB: RGB888, RGB666, RGB565, RGB555, RGB444
o YUV: YUV420-8b (legacy), YUV420-8b
o RAW: RAW6, RAW7, RAW8, RAW10, RAW12, RAW14, RAW16

which does not list 20-bit. If you wanted to follow-up in a new topic about it, one of our camera guys would be happy to double-check about the support.