Step 2: Review profiling results

Once your AI model is uploaded, you can review its detailed structure and performance profile before starting optimization.

What you can check in this step:

• Total Latency
  The model’s end-to-end inference time measured in milliseconds.

• Peak Memory
  Maximum memory consumption during inference.

• Compute Units
  Resource usage across CPU, GPU, and NPU (if applicable).

• Framework / Runtime / Graph Optimization
  Information about the model’s framework version, runtime environment, and whether graph optimization has been applied.

Layer-Level Analysis

View all layers in the model along with:

• Layer name and operation type
• Latency contribution and compute unit
• Robustness Score:
  Shows how safely a layer can be quantized without significantly affecting model accuracy. A higher score means the layer is more tolerant to precision reduction.
• Robustness Rank:
  Ranks all layers based on their robustness scores. A higher rank means the layer is more "quantization-friendly" and can be safely optimized earlier in the pipeline.
• Sorting and filtering options for precision, latency, and sensitivity

This helps you understand which layers are performance bottlenecks or accuracy-critical.

Model Visualization

On the right side, a graph view of the model architecture allows you to:

• Visually trace the computation flow
• Inspect the data shape and operator structure
• Identify repeated or heavy operations intuitively