SDAA429 June   2026 MSPM0G5187

 

  1.   1
  2.   Abstract
  3.   Trademarks
  4. 1Introduction
  5. 2MSPM0G5187 with TinyEngine NPU
  6. 3Edge AI Toolchains
    1. 3.1 TI Edge AI Studio
    2. 3.2 TI Tiny ML Tensorlab
    3. 3.3 TI Neural Network Compiler
  7. 4Edge AI Application: Digit Recognition
    1. 4.1 LeNet-5 Variant CNN Model
    2. 4.2 NPU/CPU Performance Comparison
  8. 5Edge AI Application: Waveform Classifier
    1. 5.1 Feature Extraction
    2. 5.2 Time-Series Classification Model
    3. 5.3 Model Memory Considerations
    4. 5.4 NPU/CPU Performance Comparison
  9. 6Summary
  10. 7References

4 Edge AI Application: Digit Recognition

Digit recognition technology is a fundamental capability with broad applications across embedded and personal consumer markets. However, deploying reliable recognition on embedded systems presents a significant challenge: traditional template-matching approaches lack robustness against the natural variability in human handwriting. The emergence of Edge AI offers a compelling design - enabling accurate, real-time inference directly on the device itself, without reliance on cloud connectivity. CPU-based inference pipelines demand substantial computational resources that exceed the processing capabilities typically available on resource-constrained microcontrollers.

The MSPM0G5187 microcontroller addresses this challenge by integrating a dedicated TinyEngine NPU, designed to accelerate ML workloads with minimal power and memory overhead. The design demonstrates classification of single-digit character images into one of 10 categories (0-9). The application receives a 28×28 grayscale image as pixel data over the UART back channel, performs hardware-accelerated model inference using the on-chip NPU, and sends the recognized digit class back to a host GUI for display.

Leveraging the on-chip NPU, the digit recognition system achieves approximately 99% classification accuracy with an inference latency of only 6.05ms, while consuming just 73KB of Flash and 10.9KB of RAM, as shown in Table 5-6. These results demonstrate the viability of on-device machine learning on resource-constrained microcontrollers.

Table 4-1 Digit Recognition Edge AI Design Performance
Metric Value
Accuracy Approximately 99%
Flash Usage 73KB
RAM Usage 10.9KB
Inference Latency (NPU) 6.05ms
Inference Power Consumption (AVG) 424.65uJ

Table 5-5 shows the mode main information.

Table 4-2 Digit Recognition AI Model Information
PropertyValue
Model ArchitectureCNN
Number of Parameters60,000
Input Shape(1, 28, 28)
Output Classes10
QuantizationINT8