6.2. NPU Guidelines

Some TI MCUs include a hardware neural processing unit, the TI NPU (TinyEngine NPU). This page summarizes which devices have it, how a model ends up running on it, and the key layer rules. The complete layer-by-layer constraints are in NPU Configuration Guidelines for TI Neural Network Compiler.

6.2.1. NPU Devices

`target_device`	Family	Notes
F28P55	C2000 (C28x)	Time series tasks. NPU set up by SysConfig in the C2000Ware examples.
MSPM0G5187	MSPM0 (Cortex-M0+)	Time series, image and audio tasks.
AM13E2	AM13 (Cortex-M33)	Time series and image tasks.
MSPM33C34	MSPM33C (Cortex-M33)	Has an NPU device profile, but no 1.5.0 model lists it (see Connectivity, AM26x, MSPM33 and Radar Devices).

CC2755, CC2745 and CC35X1 do not have an NPU. They accelerate the model with the Cortex-M33 Custom Datapath Extension (ti-npu type=m33cde), see Connectivity, AM26x, MSPM33 and Radar Devices.

6.2.2. How a Model Runs on the NPU

Three things must line up:

An NPU device in common.target_device.
A TINPU-quantized model. On NPU devices training.auto_quantization is on by default and produces the quantized model under training/quantization/. If the final model is not TINPU-quantized, ModelMaker compiles it with ti-npu type=soft (CPU) even on an NPU device. See Quantization.
A hardware NPU preset: default_preset or compress_npu_layer_data. forced_soft_npu_preset deliberately runs the model on the CPU.

common:
  task_type: arc_fault
  target_device: F28P55
training:
  model_name: CLS_1k_NPU
compilation:
  compile_preset_name: default_preset   # or compress_npu_layer_data to save Flash

compress_npu_layer_data adds opt_for_space=true to the compiler target, which compresses the NPU layer data. All presets are described in Device Overview.

You can tell that a compiled model uses the NPU from tvmgen_default.h: it defines TVMGEN_DEFAULT_TI_NPU (hardware NPU) or TVMGEN_DEFAULT_TI_NPU_SOFT (CPU). The SDK examples check TVMGEN_DEFAULT_TI_NPU and, when it is defined, wait for tvmgen_default_finished after tvmgen_default_run(). See NPU Device Deployment.

6.2.3. Key Layer Rules

A layer is accelerated only if it meets the NPU constraints; otherwise it falls back to software and the model gets slower. The most important rules:

Rule	Detail
Channels in multiples of 4	Input and output channels of GCONV, DWCONV, PWCONV, MAXPOOL and AVGPOOL layers must be 4, 8, 12, 16, …
First convolution (FCONV)	Input channels = 1, output channels a multiple of 4, kernel width 1 to 8.
Depthwise convolution	`groups` equal to the input channels.
Pointwise convolution	Kernel `(1, 1)`, stride `(1, 1)`.
Max pooling	At least one kernel dimension must be 1 to 4 (`(8, 1)` is valid, `(8, 8)` is not).
Global average pooling	Input must satisfy iH × iW > 2.
Fully connected	At least 16 input features with 8-bit weights, at least 8 with 4-bit weights.

The _NPU models in the ModelZoo (CLS_*_NPU, REGR_*_NPU, AD_*_NPU, FCST_*_NPU, the audio and MobileNet image models) are designed to meet these rules, and keep all kernel sizes at 7 or below. The detailed guidelines describe the cases where larger one-dimensional kernels are allowed.

6.2.4. Troubleshooting

The model runs, but slower than expected on an NPU device. Open the run’s compilation/artifacts/tvmgen_default.h. It defines TVMGEN_DEFAULT_TI_NPU when the model was compiled for the hardware NPU, and TVMGEN_DEFAULT_TI_NPU_SOFT when it was compiled for the CPU. In the second case, make sure the model is TINPU-quantized and that you did not select forced_soft_npu_preset.

Some layers are not accelerated. Compare each layer with the rules above, typically a channel count that is not a multiple of 4, a max-pooling kernel with both dimensions above 4, or fewer than 16 features into the FC layer.

The application hangs after tvmgen_default_run(). The NPU was not initialized. Keep the NPU initialization of the SDK example (see NPU Device Deployment).