Optimize the Model

With both model and calibration dataset available, you can use the eIQ AI Toolkit Optimize & Convert feature to prepare your model for deployment on NXP hardware.

What the AI Toolkit optimize&convert Does

The Optimize and convert tool runs conversion and optimization pipelines on your model:

• Applies quantization (e.g., int8)

• Performs graph-level modifications

• Convert operators to optimized kernels for NXP NPU accelerators (Neutron, Vela)

• Reduces latency and memory footprint

• Produces deployment-ready model artifacts

Using the Calibration Dataset

The calibration dataset allows the Optimizer to learn representative activation ranges for post-training quantization. Using a high-quality calibration dataset:

• Improves accuracy retention

• Reduces quantization error

• Generates more reliable profiling and benchmarking results

Optimize and Convert a Model

The Optimize & Convert page provides a visual workflow canvas where you build a pipeline by connecting your model to one or more conversion steps.

Steps:

1. Switch to the AI Toolkit tab in the top navigation bar.

2. In the left sidebar, under Model manipulation, click Optimize & Convert.

3. On the workflow canvas, your uploaded model appears as a node. If not, drag your model from the available resources onto the canvas.

4. Click the + button on the canvas to open the Select step dialog.

5. Choose a conversion step. The available steps and their compatibility depend on your model format:

   • NeutronConversion — converts a TFLite model to run on the Neutron accelerator.

   • VelaConversion — optimizes a TFLite model for NXP Ethos-U NPU.

   • OnnxConversion — converts a PyTorch model to ONNX format.

   • TFLiteConversion — converts an ONNX model to TFLite format.

   • ONNX2Quant — quantizes an ONNX model.

   Enable Show incompatible to see all steps, including those not compatible with your current model.

6. Configure the conversion parameters:

   • Target — select the target device (e.g., imxrt700).

   • Flavor (version) — select the SDK or toolchain version (e.g., MCUXpresso SDK 26.03).

   • Optional checkboxes such as Export output model as header file, Use sequencer, or Fetch constants to SRAM, depending on the conversion type.

7. Click Add to add the step to the pipeline. The step node appears on the canvas, connected to your model node.

8. Click the Run button (blue play icon) in the bottom-right corner of the canvas to start the pipeline.

9. A confirmation message appears: “Your pipeline is running…”. You can:

   • Click Optimizations History to monitor the pipeline status.

   • Click Run another to start a new pipeline.

10. When the pipeline completes, a Copying to My models dialog appears. Enter a name for the optimized model (e.g., my_model_neutronconversion) and click Copy.

11. The optimized model now appears in your model list under AI Hub. Click it to open the Model Detail page, where you can view model information, visualize the architecture, download the model file, or delete it.

Available Conversion Steps

Step	Input Format	Output Format	Description
NeutronConversion	TFLite	TFLite	Convert for Neutron accelerator
VelaConversion	TFLite	TFLite	Optimize for Ethos-U NPU
OnnxConversion	PyTorch	ONNX	Convert PyTorch to ONNX
TFLiteConversion	ONNX	TFLite	Convert ONNX to TFLite
ONNX2Quant	ONNX	ONNX	Quantize ONNX model

After optimization, the AI Hub automatically registers all generated artifacts so they can be used in downstream steps such as profiling and benchmarking.

Note

Please refer to AI Toolkit document for detailed information.

Next Steps

• Run simulated profiling