In the context of AI at the edge, STMicroelectronics has recently released the STM32N6 platform, featuring an NPU capable of 600 GOPS.
The STM32N6 is the most powerful STM32 produced by ST, with a Cortex-M55 MCU running at 800 MHz, an NPU, called Neural-ART, running at 1 GHz, a GPU, 4.2 MB of on-chip contiguous RAM, fast serial I/F for external memories, dedicated ISP and multiple camera interfaces, and many other interfaces.
The Neural-ART accelerator is the first result of a 9-year research program conducted in ST and contains 300 configurable MAC units and two 64-bit AXI memory buses, delivering a throughput of 600 GOPS, which is about 600 times more than the performances of our fastest STM32H7, which does not feature an NPU. The Neural-ART is also optimized for power consumption, with a stunning 3 TOPS/W.
The new STM32N6 is already compatible with a large number of AI operators from TensorFlow Lite, Keras and ONNX, and ST is committed to continuously increase the number of supported operators in the future.
A comprehensive ecosystem of tools
The STEdgeAI tool can be used for automatically converting pretrained neural network models to the equivalent optimized C code that can be run on one of the STM32 platforms. If the target platform is the STM32N6, this tool will create the necessary code for programming the hardware accelerators of the Neural-ART, resulting in a seamless deployment of the neural network models to the new hardware-accelerated STM32.
This solution allows the user to take advantage of the new features of the STM32N6 without going through a time-consuming learning phase.
STM32N6 in EdgeAI EU Project
STMicroelectronics is partner of the EdgeAI EU Project and is planning to use the STM32N6 board in at least one of the demonstrators of the project. The selected demonstrator is in the context of Digital Industry and regards the application of AI to quality control. It analyses the manufacts produced by a machine and identifies manufacturing defects.
The implementation of this demonstrator consists of different steps, from the acquisition of the dataset up to the deployment of the hardware and software mounted on the machine. The acquired dataset has been augmented and is being used for training.
For what concerns the NN model to be deployed on the hardware, some experiments have been conducted with non-hardware-accelerated STM32 platforms.
The same NN model has been ported to the STM32N6, obtaining performances in the range of hundreds of frames per second. Thanks to STEdgeAI tool, this port to STM32N6 was done very rapidly.
Blog signed by: STM team