AI Inference

The cube:evk is built on the i.MX8MP SoC, with a quad-core Arm Cortex-A53 @ 1.8 GHz and an integrated Neural Processing Unit (NPU) rated at up to 2.3 TOPS. The NPU accelerates machine-learning inference at the edge, reducing CPU load and enabling real-time inference under latency constraints.

The platform supports LiteRT (TensorFlow Lite) for on-device inference. LiteRT is an open-source library for executing machine learning models on mobile and embedded devices.

VX Delegate

LiteRT can execute compute operations across different compute units. The VX Delegate offloads supported operators to the on-chip hardware accelerator embedded in the i.MX 8 series.

The VX Delegate works with the hardware accelerator driver (OpenVX with extensions). It is an external TensorFlow Lite delegate, initialized by loading a dynamic library that contains the delegate implementation.

The library is located at /usr/lib/libvx_delegate.so.

Inference Example

A public example repository is available:

https://github.com/cubesys-GmbH/tflite-inference-example

The example demonstrates image detection with bounding boxes. It shows how to load TFLite models, run inference on the NPU, and visualize detections.

The cube:evk can run perception workloads such as object detection, classification, scene segmentation, or sensor fusion at the edge, without cloud dependency or external accelerators. This is relevant for V2X applications where reaction time, reliability, determinism, and privacy matter.

The cube:evk's connectivity (Ethernet, CAN, GNSS/IMU, and native DSRC / C-V2X link-layer support) lets the NPU feed AI perception results into V2X messages in real time. Typical targets are vehicles, VRUs, robots, and roadside units that need local perception combined with V2X sharing.