With rapid advancements in the machine learning (ML) space, it’s almost dizzying the array of new applications suitable for ML. At NXP, we continue to grow our ML software solutions to support the ever-expanding market. Let’s look at some of the latest technology advancements we have made available within our eIQ machine learning software environment.
All Machine Learning applications, whether for cloud, mobile, automotive or embedded, have these things in common—developers must collect and label training data, train and optimize a neural network model and deploy that model on a processing platform. At NXP, our processing platforms don’t specifically focus on cloud or mobile, but we are seriously enabling machine learning for embedded applications (industrial and IoT) and automotive applications (driver replacement, sensor fusion, driver monitoring and ADAS).
eIQ ML Software for IoT and Industrial
In June 2019, we launched our eIQ ML Software Development Environment with the primary goal of optimally deploying open source inference engines on our MCUs and application processors. Today these engines include TensorFlow Lite, Arm® NN, ONNX runtime and OpenCV, and as Figure 1 depicts, these span across all compute engines in one way or another. And wherever possible, we integrate optimizations into the inference engines (such as a performance-tuned backend for TensorFlow Lite), targeted at making our MCUs and applications processors faster. To facilitate customer deployment, we include these engines along with all necessary libraries and kernels (e.g. CMSIS-NN, Arm Compute Library) in our Yocto BSPs and MCUXpresso software development kit (SDK).
Figure 1: eIQ™ machine learning software development environment
An important part of our support for these open source inference engines is in the maintenance of version upgrades; whether they are synchronous (e.g. Arm does quarterly releases of Arm NN and Arm Compute Library) or asynchronous (Google releases TensorFlow Lite versions whenever warranted). In the fast-moving world of machine learning, these upgrades and feature enhancements are important and always deliver better performance, support for more neural network operators (to allow the use of newer models) and other new features. The release information, which is much too long to list here, is available on the github pages for Arm NN and TensorFlow.
Figure 2: eIQ Auto™ performance benchmarks
Recently, as machine learning technologies have expanded within NXP, eIQ ML software has grown to become an umbrella brand representing multiple facets of machine learning. Further enhancement of eIQ software comes from our automotive group who recently rolled out the eIQ Auto toolkit, providing an Automotive SPICE® compliant deep learning toolkit for NXP’s S32V processor family and ADAS development. This technology aligns with our S32 processors offering functional safety, supporting ISO 26262 up to ASIL-D, IEC 61508 and DO 178. The inference engine of the eIQ Auto toolkit includes a backend that automatically selects the optimum partitioning for the workload of a given neural network model across all the various compute engines in the device. The eIQ Auto toolkit also integrates functionality to quantize, prune and compress any given neural network. Benchmarks indicate that this combined process leads up to 36x greater performance for given models compared to other embedded deep learning frameworks.
Over time, we will roll out updated versions and new releases of eIQ ML software with added features and functionality to bring increased value to your machine learning applications. Without unveiling too much detail, new eIQ ML software features will include tools for model optimization (performance increase and size reduction) and enhancements to make ML software easier to use. For NXP, this is the future of machine learning—faster, smaller, easier to use software with increased functionality—all leading to widespread industry adoption.