Inference Module Real-Time Evaluation

1. Description

This report presents the real-time performance analysis of the Inference Engine module in the FluxSand project. This module utilizes a pre-trained ONNX model to classify motion or status based on time-series sensor input.

The analysis is based on 50 consecutive inference runs as recorded in the unit test log. The model takes as input a 3D tensor of shape [1, 1500, 8] and produces output of shape [-1, 10].

2. Test Environment

• Model: ONNX format, loaded from /usr/local/share/FluxSand/model/model.onnx
• Input Shape: [1, 1500, 8]
• Output Shape: [*, 10]
• Framework: ONNX Runtime
• Platform: ARM-based Linux system with RT kernel
• Thread Affinity / Priority: Not explicitly set in test, assumed normal user-space context

3. Inference Time Results (from 50 runs)

Run #	Inference Time (ms)	Result
1	14.979	Unrecognized
2	14.202	Unrecognized
3	14.242	Still
…	…	…
25	10.213	Still
33	15.805	Still
50	12.817	Still

Summary

• Total Runs: 50
• Minimum Time: 10.213 ms
• Maximum Time: 18.018 ms
• Average Time: 13.701 ms

4. Analysis

• The inference latency across 50 runs ranged from ~10.2 ms to 18.0 ms, which indicates relatively stable performance with slight variations likely due to thread scheduling or memory fluctuations.
• The average latency is 13.701 ms, allowing the system to handle inference at approximately 70–100 FPS if inference is the only load.
• The early runs exhibited slightly higher latency, likely due to initial memory paging or cache warm-up. The majority of inferences settled within 12–15 ms.
• The result output was dominantly “Still”, indicating correct environment classification consistency once the model was stabilized.

5. Conclusion

The inference module demonstrates stable and acceptable real-time performance. With average response times under 14 ms, the system is capable of responsive and interactive behavior, suitable for continuous or event-driven AI inference tasks in embedded environments.