I know that for development / compilation of networks I need an x86-64 machine running Ubuntu 16.04. However I've not found information about the requirements for actual edge devices using the NCS + a compiled deployed model for inference, e.g. in a robot or drone. Would those also have to be Ubuntu-x86-64?
@xtrawurst Yes, you are right. In order to be compatible with the NCS, the edge devices involved would have to be compatible with the x86 architecture and run Ubuntu 16.04 64-bit. X86 boards like the Up squared board and Minnow board should work fine. In addition, the Raspberry Pi is another option worth considering. Inference is supported on the RPi, although compilation and development will need to be performed on a laptop or computer.
Thanks. How is it possible that the RPi is supported for inference given that it's running on an ARM chip (i.e. no x86)?
Another question: Is it possible to use the NCS for general image processing tasks (i.e. things the Myriad 2 has been designed for such as feature detection and tracking, warping, convolutions etc..), not just for running neural networks?
@xtrawurst Support for RPi3 was added in MvNC_API-1.07.07. Previous Toolkit version only supported x86-64 machines for both learning and inference. Now you can use either x86-64 or RPi3 as host for NCS with previously profiled and compiled network on a Desktop/Laptop x86-64.
@xtrawurst Although it uses the versatile Myriad2, the NCS product is designed to perform inferences using neural networks on edge devices so it isn't possible to use the NCS for other general image processing tasks.
Thanks. Since the NCS is not suitable for general vision / image processing tasks, is there another product / devkit by Movidius available today for developers that uses the Myriad 2?
The Myriad Development Kit & Board (https://www.movidius.com/myriad2) looks interesting but it's unclear to me how to obtain one.
@xtrawurst No problem. Please keep in mind that the NCS is designed to accelerate neural network inferences on the edge. General image processing and computer vision pre-processing should be done on the host platform and processing related to neural network inferences can be offloaded to the NCS. If you are interested in edge devices running the NCS, other forum users have made threads about running the NCS on edge devices like the raspberry pi and documented the results. Please feel free to check it out at: https://ncsforum.movidius.com/discussion/146/raspberry-pi-3-benchmarks#latest