11. Release Notes

Software Manifest

TI_MCU_2.1.0

  • TI_MCU_2.1.0 User’s Guide

  • Supported devices

  • Supported all layer configs with 8-bit activations and 8-/4-/2-bit weights that can be offloaded to TI-NPU

  • Supported all layer configs with 8-bit activations and 8-bit weights that can be accelerated using the M33 Custom Datapath Extension (CDE).

  • Improved quantized (8-bit) model inference performance on C29x core. Note that floating-point models typically run faster than their quantized counterparts on C29x core, despite requiring a larger memory footprint.

TI_MCU_2.0.0

  • TI_MCU_2.0.0 User’s Guide

  • Supported devices

  • Expanded layer configs that can be supported on TI-NPU

  • Added an option to compress TI-NPU layer data

  • Optimized floating point inference on C29x core

  • Supported quantized models in QDQ format with integer inference code on CPU

TI_MCU_1.3.0

  • TI_MCU_1.3.0 User’s Guide

  • Windows and Linux support

  • Multiple model support in the same application

  • Generic padding support

  • Optimized performance for motor fault classification models and arc fault detection models

    • With TI-NPU acceleration, these models saw a 1.2x to 1.5x speedup from TI_MCU_1.2.0 release

    • With software-only execution on C28, these models saw a 3.9x to 4.8x speedup from TI_MCU_1.2.0 release

  • Added example and documentation on how to apply TI-NPU quantization to an existing PyTorch training script

  • Simplified command line options

TI_MCU_1.2.0