STDA062 October   2026 TDA4VH-Q1

 

  1.   1
  2.   Abstract
  3.   Trademarks
  4. 1Introduction
  5. 2SoC Architecture
    1. 2.1 Vector DSP and Matrix Math Compute Cluster
    2. 2.2 Vision and Motion Processing Accelerators
    3. 2.3 ARM® Cortex®-A72 CPU Cores
    4. 2.4 ARM Cortex-R5F CPU Cores
    5. 2.5 On-Chip Memory and High-Performance Interfaces
    6. 2.6 Package and Power Consumption
  6. 3Machine Learning Software Development Tools
    1. 3.1 TI Deep Learning Tools
    2. 3.2 TI Edge AI Studio
  7. 4TDAx Family Scalability
  8. 5Conclusion
  9. 6References

3.1 TI Deep Learning Tools

From a machine learning software development standpoint, the TI Deep Learning (TIDL) tools [5] use a bring-your-own-model (BYOM) approach, making it possible for data scientists, or embedded system designers, to reuse existing custom tool flows or open-source AI training platforms. The following industry-standard machine learning model runtime format options are available for deploying neural network models on the SoC:

  • TensorFlow Lite-based inference
  • ONNX Runtime-based inference
  • TVM/Neo-AI-DLR based inference

The TIDL library allows each of these runtime models to be compiled for execution exclusively on C7 NPU (TIDL-RT) or with heterogeneous execution across Cortex-A72 + C7 NPU (OSRT).