Jetson Jetpack - Which GPU Architecture?

I’m new to NIVIDIA and this forum, so apologies if this question has been answered. I did look!

I’m developing an advanced AI/ML class for my university and the Jetson looks like the perfect compute device for our labs.

I want to introduce the students to Mixed Precision Floating Pt. Math to optimize their models. I see that this is supported on the NVIDIA Turning and Volta architecture. Turning adds this mixed casting with integers.

What GPU architecture is in the Jetson? I know it has CUDA and Tensor Cores, so from a hardware POV that’s promising.

And my students use Scikit-Learn, but this platform isn’t supported by NVIDIA. I found an unsupported way to make it work with Tensorflow calls, but not lower-level than that. Will this ever be supported by NVIDIA?

I see Tensorflow and PyTorch is supported by NVIDIA’s Mixed FP Precision libraries, so I may have a way to make it work for SciKit-Learn I hope.

Anyone experienced with NVIDIA AMP (Automatic Mixed Precision) who can advise me on whether to use it, or make the students manually adjust their precisions, normalizations, etc? I will teach them what’s happening either way.

I assume I should install the latest JetPack 6,0 software onto the Jetson Orin AGX 64 I plan to purchase, True?

Thank you!

Francesco Bonifazi - University of Colorado - Denver


Jetson has lots of series and cross-generation. You can find more details here.
Orin series (including AGX, NX, and Nano) are Ampere, and the Xavier series is Volta.

scikit-learn can be installed on Jetson and should work normally.
TensorFlow and PyTorch also can work on Jetson correctly.

AMP is usually used for training.
On Jetson, we usually use post-training quantization with the original fp32 model.
This can be done by converting the model with TensorRT.