New Milestone: emx-onnx-cgen on the ONNX Backend Scoreboard

emx-onnx-cgen is now officially listed as an ONNX backend on the ONNX website.

The ONNX Backend Scoreboard evaluates implementations against the official ONNX test suite –providing a transparent and standardized measure of compliance.

Our current result: ~98% test pass rate, putting emx-onnx-cgen roughly on par with the ONNX reference implementation.

What is emx-onnx-cgen?
The emmtrix ONNX-to-C Code Generator (emx-onnx-cgen) is an open-source AI frontend compiler that translates ONNX models into deterministic, analyzable C code. The generated code is hardware-agnostic and designed for transparency, determinism, and static analyzability – key requirements for safety-critical and resource-constrained systems.

In combination with the emmtrix Vectorizer, this enables resource-optimized deployment of AI models on embedded architectures such as Infineon Technologies TC4x PPU, RISC-V (RVV), and ARM NEON/SVE. See our wiki article.

How we got here
• v1.0 – Full ONNX Opset 25 operator coverage
• v1.1 – Completion of all ONNX data types
→ including 2-bit and 4-bit integers, as well as 4-bit and 8-bit floating point formats
→ achieving full ONNX Opset 25 test coverage
• v1.2 – Introduction of an ONNX backend interface
→ using the existing C code generation pipeline under the hood

This means emx-onnx-cgen is no longer just a code generator, but also a fully testable ONNX backend with measurable compliance.

Why this matters
• Independent validation against the official ONNX test suite
• Direct comparability with other ONNX runtimes and backends
• Increased confidence for real-world deployment scenarios
• Strong indication of correctness and specification adherence

This milestone demonstrates that generating portable C code from ONNX models can achieve a level of correctness comparable to established runtimes.