9.1.3. NPU Device Deployment

This page lists what is specific to deploying on devices with a hardware NPU. The common procedure (the four files, picking the SDK example, the expected console output) is in Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check); the CCS steps are in CCS Integration Guide.

9.1.3.1. NPU-Enabled Devices

`target_device`	SDK example location	NPU initialization
F28P55	C2000Ware `libraries/ai/examples/<example>/f28p55x/`	SysConfig (NPU module in `c2000.syscfg`)
MSPM0G5187	MSPM0 SDK (not part of `tinyml-sdk`; see MSPM0 Family)	See the MSPM0 SDK example
AM13E2	MCU-SDK `examples/ai/<example>/`	See the MCU-SDK example

MSPM33C34 has an NPU profile but no model lists it in 1.5.0 (Device Overview). F29 devices (F29H85, F29P58, F29P32) have no hardware NPU: the model is compiled with the soft NPU preset and runs on the CPU through TI’s NPU software library. That is why the F29H85 examples still call TI_NPU_init() in application_main.c and carry an NPU module in their SysConfig; do not remove either from an example you import.

9.1.3.2. NPU Compilation

An NPU device is compiled with one of the NPU presets (default_preset, compress_npu_layer_data); forced_soft_npu_preset deliberately runs the model on the CPU. Model layers must satisfy the NPU rules in NPU Guidelines and NPU Configuration Guidelines for TI Neural Network Compiler; models named *_NPU in the model list are designed for that. The result is the same four files as for any device.

common:
  target_device: 'F28P55'

compilation:
  enable: True
  compile_preset_name: default_preset   # or compress_npu_layer_data to save Flash

9.1.3.3. What Is Different at Run Time

Initialization. The NPU must be initialized before the first tvmgen_default_run. In the SDK examples this is already done (SysConfig, or TI_NPU_init() for F29H85). Keep the NPU module in the project’s .syscfg.
Asynchronous completion. On NPU devices the application waits for tvmgen_default_finished before it compares the output. If the program hangs there, the NPU was not initialized.
Tolerance. The golden-vector check uses the same tolerance (+/-1, or +/-4 for anomaly detection) as on CPU devices.
Memory. Check the linker map after the build. ModelMaker also reports model and feature-extraction memory (see Device Overview).

9.1.3.4. Walkthrough: F28P55x

Install C2000Ware and CCS (CCS Integration Guide).
Import f28p55x_arc_fault (or the example for your task) from Resource Explorer using the board filter LAUNCHXL-F28P55X.
Replace artifacts/mod.a, artifacts/tvmgen_default.h, test_vector.c and user_input_config.h with the files from one NPU-compiled run (for example examples/dc_arc_fault/config_dsk.yaml with model CLS_1k_NPU).
Build, select TMS320F28P550SJ9_LaunchPad.ccxml, debug, add test_result to Watch and resume.
Expect Feature extraction mismatches 0 and All tests have passed..

9.1.3.5. Troubleshooting

Symptom	Fix
Hangs at `while (!tvmgen_default_finished)`	NPU not initialized. Keep the NPU module in `.syscfg` (C2000) or `TI_NPU_init()` (F29H85).
Linker: incompatible object / undefined `tvmgen_default_run`	`mod.a` was built for a different device. Re-run with the right `target_device`.
Model compiled but not on the NPU	A layer is not NPU-compatible, or `forced_soft_npu_preset` was used. See NPU Guidelines.
Section does not fit in Flash	Try `compress_npu_layer_data` or a smaller model.