Edge AI Using MATLAB And Simulink

Machine learning and deep learning models can be simulated, tested, and deployed on edge devices and embedded systems. Create code for full AI applications, including pre- and post-processing algorithms.

You can use MATLAB® and Simulink® to:

Create optimised C/C++ and CUDA code for use on CPUs and GPUs.
Create synthesizable Verilog and VHDL code for FPGAs and SoCs.
Accelerate inference with deep learning libraries optimised for hardware, such as oneDNN, Arm Compute Library, and TensorRT.
Incorporate pre-trained TensorFlow Lite (TFLite) models into hardware-deployed applications.
AI models for inference on resource-constrained hardware can be compressed using techniques for hyperparameter tuning, quantization, and network pruning.

Microcontrollers and CPUs
MATLAB CoderTM and Simulink CoderTM generate portable, optimised C/C++ code from taught machine learning and deep learning models. In the produced code, you can incorporate calls to vendor-specific libraries optimised for deep learning inference, such as oneDNN and the Arm® Compute Library.

GPUs
GPU CoderTM generates optimised CUDA® code for deep learning networks. To distribute entire algorithms to desktops, servers, and embedded GPUs, include pre- and post-processing in your networks. To improve speed, use NVIDIA® CUDA libraries such as TensorRTTM and cuDNN.

SoCs and FPGAs
The Deep Learning HDL ToolboxTM allows you to prototype and implement deep learning networks on FPGAs and SoCs. Deep learning processors and data movement IP cores can be programmed using pre-built bitstreams for popular FPGA development kits. HDL CoderTM can be used to create custom deep learning processor IP cores and bitstreams.

Compression of AI Models
Hyperparameter tuning and quantization of weights, biases, and activations reduce memory needs for machine learning and deep learning models. Reduce the size of a deep neural network by cutting unnecessary layer connections.