Qualcomm promised integration with Google’s TensorFlow for onboard machine learning processes in the Snapdragon 835, and with today’s official release of the Snapdragon Neural Processing Engine SDK, it delivered on that and then some. For starters, TensorFlow is not the only engine supported – you can also port in work from Caffe and Caffe2. On top of that, applications integrating the functionality can be written in Java or native formats for Android or Linux. The SDK allows a trained AI model to be run directly on board a Qualcomm Snapdragon 835 device without the need to connect to the cloud, and allows developers to integrate such models into almost any application.
Onboard neural networking means that onboard AI operations don’t have to depend on a good network connection, though the Snapdragon 835 does boast gigabit LTE connectivity. In any case, this means that the powerful chip is able to run virtual machines for neural network purposes, and is capable of onboard, independent machine learning when handed a trained model, so long as the application that the stack is integrated into calls for such. Things like offline AI assistants, powerful and smart AI enemies in games, and on-device AI-driven analytics for things like malware protection and user data crunching are all possible with this SDK. The best part is that a trained model made in TensorFlow or Caffe only has to be converted into the Snapdragon-friendly DLC format, with no additional tweaking, in order to be run and integrated on-device.
The SDK comes with sample code, heavy documentation, performance monitoring and bottleneck identification mechanisms, and an easy tool for converting premade models to the right format, then shoehorning them into an application. To put it as simply as possible, at this point, anybody with an Ubuntu 14.04 or higher compatible PC and a Qualcomm Snapdragon 835 device lying around can take advantage of the vast possibilities and power of neural networking and machine learning. The SDK supports many different kinds of neural networks and machine learning scenarios, including some of the most common like convolutional neural networks, and can even integrate with stacks for things like IoT and automotive devices. Qualcomm is also planning on implementing more different model creation tools later on. If you have an Android or Linux app and an AI model built in TensorFlow or Caffe, and want to integrate them, head through the source link.
