Andrew Su is a senior technical expert in the AI product direction at Renesas Electronics China, responsible for product definition and operation of AI-related products and technical solutions for the Chinese market.
What is RUHMI?
With the launch of the new high-performance RA8P1 microcontroller integrated with the Ethos-U55 NPU and Arm Cortex-M85 CPU core, Renesas provides a set of tools for deploying common AI models onto high-performance computing platforms to fully utilize the hardware AI computing acceleration units of the target platform for AI computation tasks, which is RUHMI.
RUHMI stands for “Robust Unified Heterogeneous Model Integration,” meaning a robust, universal, adaptable model converter for heterogeneous chip platforms. As shown in Figure 1.
Figure 1: Full name of RUHMI
Using the RUHMI toolkit, it is possible to convert common machine learning development platforms, such as models developed with TensorFlow, PyTorch, and saved in ONNX format, into C source code (executed by the CPU) and binary executable programs (executed by the NPU) that can be directly deployed on Renesas processor platforms (e.g., RA8P1).
Especially for hardware platforms integrated with dedicated NPUs like the Ethos-U55, the execution efficiency of machine learning models deployed using the RUHMI toolkit is significantly higher than that achieved using general-purpose tools for the same functionality.
Developers familiar with Renesas’ AI product family may recall the DRP-AI TVM tool developed to support Renesas’ self-developed DRP-AI hardware acceleration engine, which can deploy some visual AI models onto the RZ/V chip platform (Arm Cortex-A55 + DRP-AI). In fact, the design of RUHMI draws on key technologies from DRP-AI TVM.
RUHMI provides developers at different stages with both command line and graphical interface options to offer users a better AI compiler/converter development experience. As RUHMI continues to support more Renesas chip products, it will integrate the AI development toolchain for MCUs and MPUs, separating hardware differences from model development, allowing AI application developers to focus more on model design and development while benefiting from optimized execution efficiency at the hardware level.
Technical Details of RUHMI
RUHMI is a toolkit that contains a series of tools capable of performing tasks independently. For example, RUHMI uses a model parser (Model Importer/Parser) to convert various open-source model structure files into intermediate layer code (Intermediate Representation), then optimizes the model through a model optimizer (Model Optimizer), including quantization (converting FT32 to INT8), fine-tuning, and pruning, and then separates the computational processes that can be handled by hardware acceleration units (NPU) from those that require software processing (CPU), generating source code or executable binary files suitable for dedicated hardware units and CPU.
Figure 2: Functional block diagram of RUHMI
Getting Started with RUHMI Development
Using the RUHMI development flow is very simple. For developers, first visit the RUHMI product homepage of Renesas to access the RUHMI project homepage on GitHub, then download the corresponding tool software package and install it on the local host, run the tool to convert, and then integrate the converted code/binary files into the embedded project source code, compile and download to complete the deployment. As shown in Figure 3.
Figure 3: Development flow using RUHMI
Among them, the RUHMI project homepage on GitHub is shown in Figures 4-5.
👈 Swipe left and right to view Figures 4-5 👉
RUHMI project homepage on GitHub
Scan the QR code below or copy the link to view related information.
RUHMI project homepage on GitHub
https://github.com/renesas/ruhmi-framework-mcu
There are two ways to convert models using RUHMI: command line and graphical user interface. The graphical interface is suitable for simple trials of the model converter, while for experienced developers, it is recommended to use the command line for access to more functional APIs and easier integration into automated workflows. Additionally, the command line development method is also applicable to Linux system platforms. As shown in Figure 6.
Figure 6: RUHMI command line and graphical user interface
The graphical user interface tool of RUHMI is integrated into Renesas’ e2Studio integrated development environment, and the generated source files and binary files are ultimately compiled in the e2Studio environment to generate firmware files deployed on the target chip, which can be downloaded to run on the chip.
Examples and Performance
In fact, the first released RUHMI toolkit already includes two use cases, based on Renesas’ official EK-RA8P1 circuit board, implementing face recognition and image classification models deployed on the RA8P1 chip. The system block diagram of the design program is shown in Figure 7.
Figure 7: Two example projects of RUHMI: face recognition and image classification
By running these two use cases separately, with the NPU (Ethos-U55) enabled, the inference speedup compared to not using the NPU is nearly 20 times. As shown in Figure 8.
Figure 8: Significant speedup of RUHMI using NPU
Examples and Performance
Currently, the RUHMI product homepage on GitHub is open to the public, providing ample documentation, software tools, and example projects for developers to download, install, and try out.
Renesas Electronics will organize online and offline training and technical exchanges for developers in China.
Need Technical Support?
If you have any questions while using Renesas MCU/MPU products, you can scan the QR code below or copy the URL into your browser to access the Renesas Technical Forum for answers or online technical support.
https://community-ja.renesas.com/zh/forums-groups/mcu-mpu/
1
END
1
Recommended Reading
New Edge AI Revolution RA8P1 AI MCU: 1GHz CM85 CPU + 0.25TOPS Computing Power NPU
Zero-threshold hands-on NVIDIA TAO! Renesas AI model deployment tool demonstration
elexcon 2025 | Renesas collaborates with ecosystem partners to present AI + Industrial Technology Awards
