6. On-board implementation

6.1 Various materials

As the operations performed by neural networks can be parallelized, GPUs or FPGAs can be used to accelerate calculations considerably.

Since the mid-2010s, embedded-compatible gas pedals have appeared, packing hundreds of GPUs into a stick with a format similar to that of a USB stick. These include :

• Intel Movidius NCS stick ;

• Google Coral USB stick ;

• Nvidia Jetson Nano.

The computational load is much more intense during training than during inference. Networks are therefore generally trained on servers or large GPUs before being implemented on embedded hardware. Inference uses real-life data,...