What processor is reccommended for working with the Jetson AGX Xavier


I am working with the Jetson AGX Xavier and I was wondering what CPU is is recommended for working with this? Currently, I am looking into Intel i7 processor. Would this work?

If you mean just flashing, then any CPU will work. I even have a little dual core Atom that works (embarrassingly slow and has to remote display to have a big enough screen). If you mean actually developing, then even a 6-core CPU of most any time is rather fast as long as you have the right GPU for any CUDA development you are interested in. If you are looking into training models, then you still depend on the GPU for much, but more RAM and more cores would probably be a good idea to speed up training. Just depends on what you are doing. I have no doubt that an i7 would be good for many uses. Someone else may have suggestions, but you would need to provide more details on just what it is you want to do.

Hi @linuxdev I am interested in using the Jetson AGX Xaiver to identify text in live video and translate it to text format (likely using tesseract). I am looking to apply transfer learning using an existing CNN model to train my model.

To flash and to put pretrained models on the Xavier has almost no requirement other than being an Ubuntu 18.04 host PC. For training I think you’ll find you need a GPU with lots of RAM, and this is one of the reasons why people use Quadro or Titan or other “high GPU memory capacity” video cards. I have no doubt that any of the current high end gaming cards could also do well with training, but you might run out of memory (in the GPU). Someone else with more experience in training and development using CNN can probably add to the conversation.

Would the GPU on the Jetson AGX Xavier be sufficient?

The Xavier models are quite good at using a pre-trained model. None of the embedded systems are very good with training. If you wish to train, then you could in theory use the Xavier, but you wouldn’t be very happy with performance. You’d want to use a PC with a discrete GPU for that, or one of the cloud solutions if it is a really big model.

1 Like

Ahh, that makes sense. Thank you!