Arm NN enables third parties to run neural network models on their custom inference hardware. It also allows the addition of new backends through the new ‘Pluggable Backend’ mechanism. Arm NN's architecture has been extended to make a new backend API available to developers. They want to take advantage of Arm NN's high-level optimization and scheduling capabilities on their custom accelerated hardware. Developers can plug new backends into an Arm NN build and instruct Arm NN's optimization. They can leverage them at runtime by selecting their new backends as a valid resource for running inference workloads. The developers will also have control on which order the backends should be chosen for inference, or which fallback to resort to in case the execution of a particular part of the network is not possible or supported by the chosen hardware. To get the most out of their dedicated hardware, third parties have the possibility to apply their own custom optimizations on the original neural network model, resulting in optimal performance. Arm NN takes care of the rest.
The guide available provides detailed instructions on how to add your own specialized backend to Arm NN, and it provides a working example to help you getting started.
Access the ML platform