My name is Krishna. I intend to use the Jetson AGX Xavier dev kit to train and deploy 3D Object Detection models.
My first question is that, can neural network models be trained on the Jetson AGX Xavier from scratch just like we can on VMs?
My second question is that, can Jetson AGX Xavier be used like a Raspberry Pi (like an independent computer) without the need of a host device. Is the host device needed all the time to use Jetson or just initially to install Ubuntu onto the Jetson AGX Xavier?
The answers to these questions are really helpful for my project. Please let me know.
Xaviers can train, but the performance will be terrible. The embedded systems are really meant to run a pre-trained model.
The flash software for Xavier is run from a Linux host PC (VMs can be made to work but are generally prone to problems and a learning curve). Once flashed the Xavier does not require any other system, it is a full Linux computer.
Thanks a lot for the clarification. Do you suggest any cheap alternatives to costly Azure VMs, AWS or GCP Deep Learning VMs for training Deep Learning models?
Any Linux PC (especially easy for Ubuntu) with a desktop GPU from NVIDIA can do this. It really depends on what you are doing as to how much/long training will take, and I wouldn’t be able to give any good advice on that. Generally speaking, if you are going to train on your host PC, then you want something with more VRAM, e.g., a 6GB video card will be very limited, an 8GB somewhat limited, but once you get to 11 or 12GB your options will go up. A big reason people are willing to pay more for the Titan products is the combination of much more VRAM, and faster VRAM.
If you give more details on what you need to train, then someone else can probably give a better answer than I can.