ONNX Operators Support List
SiMa MLSoC supports models from various frameworks, provided they can be converted to ONNX (versions 16 or 17) or TFLite (version 2.10.0). ONNX is the primary format for structured verification, though some models—such as those using Caffe operators—can still run on the device, even if ONNX lacks direct equivalents. Additionally, pre-quantized TFLite models are supported without requiring further optimization.
The ModelSDK is designed to maximize performance by leveraging hardware acceleration for many ML operations. While some features are fully supported with acceleration, others are supported only without acceleration. If a feature in a model is unsupported, for example dynamic sized tensors, then the entire model is unsupported. However, if certain operations are simply not accelerated, the rest of the model can still benefit from acceleration where supported. The SDK also performs automatic optimizations, such as constant folding, to eliminate unsupported operations when possible. As a result, models that initially appear unsupported may still be successfully processed. If issues arise, it’s recommended to revise the model to align with supported features or use the graph surgery API to remove unsupported elements. Testing model quantization is an effective way to confirm support and optimize model compatibility.
The table below outlines acceleration support for each operator but does not include operators that are only supported without acceleration.
SiMa.ai is continuously working to expand acceleration support for additional operators and improve compatibility. For the latest updates on newly supported ONNX and TFLite features, be sure to check the release notes.
| Operator | 5D Tensor | Constraints | INT8 | INT16 | BFLOAT16 |
|---|
Note
BFLOAT16 support is exclusive to the next-generation Modalix platform, which is currently available through the developer preview program. For more details, reach out to support@sima.ai.
Certain operators listed in the table below are compatible with 5D tensors, structured as (batch size, depth, height, width, channels) or (N, D, H, W, C).