Integrated in MCUXpresso and Yocto development environments, eIQ delivers TensorFlow Lite for NXP’s MCU and MPU platforms.
Developed by Google to provide reduced implementations of TensorFlow (TF) models, TF Lite uses many techniques for achieving low latency such as pre-fused activations and quantized kernels that allow smaller and (potentially) faster models. Furthermore, like TensorFlow, TF Lite utilizes the Eigen library to accelerate matrix and vector arithmetic.
TF Lite defines a model file format, based on FlatBuffers. Unlike TF’s protocol buffers, FlatBuffers have a memory footprint an order of magnitude smaller allowing better use of cache lines, leading to faster execution on NXP devices. Although TF Lite supports a subset of TF neural network operations, it still maintains a wide range of supported operations (~50), thereby allowing the use of more advanced models compared to CMSIS-NN. It also supports recurrent neural networks (RNNs) and the even more complex long short-term memory (LSTM) cells.