Overview on the Renesas RZ/G3E NPU solution

• Applications running in Linux can trigger inference to be run on the Arm Ethos-U55 NPU and Arm Cortex-M33 core.
  • Operators that are unable to be run on the NPU will be offloaded to the Cortex-M33 during compilation stage.
• Prebuilt binaries available to speed-up evaluation.
• OpenEmbedded/Yocto layer provided that adds supports for the Arm Linux Driver Stack v24.08, EdgeCortix MERA runtime library and sample applications.
  • This is based on the Renesas RZG3E-BSP-1.0.0.
• Docker Environment provided with functionality to automatically generate runtime data for all provided target models.
  • The Renesas RUHMI Framework AI Compiler is used to compile AI models into data used by the MERA runtime library.
  • MERA runtime library is used to interact with the Ethos-U libraries to run an AI inference.
  • Detailed information about model compilation can be found HERE.

The diagram below illustrates the architecture of the RZ/G3E Ethos-U software stack:

Package Architecture

Supported Models

Optimized networks generated using the RUHMI Framework AI Compiler, with a file size up to 16MB.