6.1. Device Overview

This page lists every common.target_device value that Tiny ML ModelZoo 1.5.0 accepts, with the facts you need to choose one: CPU core, hardware NPU, supported tasks, whether a ready-made SDK example exists, and whether feature-extraction memory is estimated.

Note

The tables on this page are generated from the 1.5.0 source: the TARGET_DEVICE_* constants and device compilation profiles in ModelMaker (tinyml_modelmaker/ai_modules/{timeseries,vision,audio,radar}/constants.py) and the target_devices of each model in tinyml_modelzoo/model_descriptions. For clock speed, memory size and peripherals, see the device’s product page on ti.com.

6.1.1. Selecting the Device

Set the device in the common section of the config YAML. Use the exact string from the tables below (it is case sensitive):

common:
  task_type: generic_timeseries_classification
  target_device: F28P55

ModelMaker checks the combination before it starts. If the device is unknown, or no model for the task_type supports the device, the run stops with an error that lists the supported devices or tasks.

6.1.2. Complete Device List

Task legend: C generic time series classification (generic_timeseries_classification), Arc arc fault, Motor motor fault, Blower blower imbalance, ECG ECG classification, PIR PIR detection, R regression, F forecasting, AD anomaly detection, Img image classification, Aud audio classification, Rad radar classification.

SDK example names the device folder of the ready-made examples in the TI SDKs (see Deployment). FE memory means ModelMaker compiles the SDK feature-extraction library for this device and reports its memory.

6.1.2.1. C2000 F28x (C28x DSP)

Device	Core (compiler flags)	NPU	Tasks	SDK example	FE memory
F28P55	C28x, FPU32	Yes	C, Arc, Motor, Blower, R, F, AD	C2000Ware f28p55x	Yes
F28P65	C28x, FPU64	No	C, Arc, Motor, Blower, R, F, AD	C2000Ware f28p65x	Yes
F28P551x	C28x, FPU32	No	C, Arc, Motor, Blower, ECG, R, F, AD	–	No
F2837	C28x, FPU32	No	C, Arc, Motor, Blower, ECG, R, F, AD	–	Yes
F2837xS	C28x, FPU32	No	C, Arc, Motor, Blower, ECG	–	No
F2838x	C28x, FPU32	No	C, Arc, Motor, Blower, ECG, R, F, AD	–	No
F2807x	C28x, FPU32	No	C, Arc, Motor, Blower, ECG	–	No
F28003	C28x, FPU32	No	C, Arc, Motor, Blower, ECG, R, F, AD	C2000Ware f28003x	Yes
F28004	C28x, FPU32	No	C, Arc, Motor, Blower, ECG, R, F, AD	C2000Ware f28004x	Yes
F28002x	C28x, FPU32	No	Motor, Blower, ECG, R, F, AD	–	No
F280013	C28x, FPU32	No	Arc, Motor, Blower, ECG, R, F, AD	–	Yes
F280015	C28x, FPU32	No	Arc, Motor, Blower, ECG, R, F, AD	–	Yes
F28E12	C28x, FPU32	No	R, F, AD	–	Yes

6.1.2.2. C2000 F29x (C29x DSP)

Device	Core (compiler flags)	NPU	Tasks	SDK example	FE memory
F29H85	C29x	No	C, Arc, Motor, Blower, R, F, AD	F29H85x SDK f29h85x	Yes
F29P58	C29x	No	C, R, F, AD	–	No
F29P32	C29x	No	C, R, F, AD	–	No

6.1.2.3. Arm Cortex-M0+, Cortex-M33 and Cortex-R5F MCUs

Device	Core (compiler flags)	NPU	Tasks	SDK example	FE memory
MSPM0G3507	Cortex-M0+, no FPU	No	C, Arc, Motor, ECG, PIR, R, F, AD, Img	MSPM0 SDK (not in this flow, see MSPM0 Family)	Yes
MSPM0G3519	Cortex-M0+, no FPU	No	C, Arc, Motor, ECG, PIR, R, F, AD, Img	MSPM0 SDK (see MSPM0 Family)	No
MSPM0G5187	Cortex-M0+, no FPU	Yes	C, Arc, Motor, ECG, PIR, R, F, AD, Img, Aud	MSPM0 SDK (see MSPM0 Family)	Yes
MSPM33C32	Cortex-M33, FPv5-SP	No	C, Arc, Motor, Blower, PIR, R, F, AD	MSPM33 SDK mspm33c321a	No
MSPM33C34	Cortex-M33, FPv5-SP	Yes (profile only)	None in 1.5.0 (no model lists this device)	–	No
AM13E2	Cortex-M33, FPv5-SP	Yes	C, Arc, Motor, Blower, R, F, AD, Img	MCU-SDK (AM13E2x)	Yes
AM263	Cortex-R5F, VFPv3-D16	No	C, Arc, Motor, Blower, R, F, AD	MCU-PLUS-SDK am263x_lp	No
AM263P	Cortex-R5F, VFPv3-D16	No	C, R, F, AD	–	No
AM261	Cortex-R5F, VFPv3-D16	No	C, R, F, AD	–	No

6.1.2.4. Connectivity and Radar Devices

Device	Core (compiler flags)	NPU	Tasks	SDK example	FE memory
CC2755	Cortex-M33 with CDE	No	C, Motor, PIR	–	No
CC2745	Cortex-M33 with CDE	No	C, Motor, PIR	–	No
CC35X1	Cortex-M33 with CDE	No	C, Motor, PIR	–	No
CC1352	Cortex-M4F	No	C, Motor, PIR	–	No
CC1312	Cortex-M4F	No	C, Motor, PIR	–	No
CC1354	Cortex-M33, FPv5-SP	No	C, Motor, PIR	–	No
CC1314	Cortex-M33, FPv5-SP	No	C, Motor, PIR	–	No
IWRL6432	Cortex-M4F	No	Rad	–	No

Note

• MSPM33C34 has an NPU compilation profile and is accepted on the command line (it is hidden from the Model Composer GUI), but no 1.5.0 model lists it, so every task is rejected for it. Use MSPM33C32 for MSPM33 projects.

• CDE devices (CC2755, CC2745, CC35X1) have no NPU. Their default preset compiles with ti-npu type=m33cde, which uses the Cortex-M33 Custom Datapath Extension. See Connectivity, AM26x, MSPM33 and Radar Devices.

• Device-specific models exist too, for example WiFiPresDet_4.7k_CDE (CC35X1 only) and the audio models (MSPM0G5187 only).

6.1.3. Compilation Presets

ModelMaker compiles the trained ONNX model with the TI MCU Neural Network Compiler. The compile options come from a preset, selected with compilation.compile_preset_name:

compilation:
  enable: True
  compile_preset_name: default_preset   # or compress_npu_layer_data, forced_soft_npu_preset

Preset	Available on	What it does
default_preset	All devices	Recommended. Uses the hardware NPU when the device has one; otherwise compiles for the CPU. Options are chosen per task (for example, regression and forecasting do not request integer output).
compress_npu_layer_data	F28P55, MSPM0G5187, AM13E2 (and MSPM33C34, see note above)	Hardware NPU, plus opt_for_space=true: compresses the NPU layer data to reduce Flash use.
forced_soft_npu_preset	F28P55, MSPM0G5187, AM13E2, CC2755, CC2745, CC35X1	Disables the hardware acceleration and compiles the model for the CPU (ti-npu type=soft).

If you ask for a preset that the device does not have, ModelMaker logs a warning and uses default_preset.

Important

The hardware NPU only runs TINPU-quantized models. If the model is not TINPU-quantized (training.quantization is not 2), ModelMaker replaces ti-npu type=hard with ti-npu type=soft and the model runs on the CPU, even on an NPU device. training.auto_quantization is on by default and is switched off automatically on devices without a hardware NPU (unless you set it in the config). See Quantization.

6.1.4. NPU and Non-NPU Devices

Hardware NPU devices: F28P55, MSPM0G5187, AM13E2. The TI NPU accelerates convolution, pooling and fully connected layers of quantized models. To be accelerated, a layer must meet the NPU constraints (channel multiples of 4, kernel and pooling limits, minimum FC input size). Layers that do not meet them fall back to software. See NPU Guidelines.

All other devices run the model on the CPU. The _NPU models in the ModelZoo (for example CLS_1k_NPU) are not limited to NPU devices: they list CPU-only devices in their target_devices as well, and ModelMaker compiles them for the CPU there. Tiny ML ModelZoo 1.5.0 has no separate “non-NPU” variant of these models.

6.1.5. Feature-Extraction Memory Estimation

New in 1.5.0: for the devices marked FE memory: Yes above, the compilation step also compiles the SDK feature-extraction library with the run’s user_input_config.h and reports its memory in the log:

============================================================
TI Feature Extraction Library Memory Usage
============================================================
Code:                      ...
RO Data:                   ...
RW Data:                   ...

The devices are AM13E2, MSPM0G3507, MSPM0G5187, F29H85, F28P55, F28P65, F2837, F28003, F28004, F280013, F280015 and F28E12. You can also estimate model and feature-extraction memory before training with the standalone script in the tinyml-modelzoo folder:

./estimate_memory.sh examples/dc_arc_fault/config_dsk.yaml --target_device F28P55

See Memory Estimation for details.

6.1.6. Choosing a Device

• Start from the task. Check the Tasks column. Image classification runs on MSPM0 and AM13E2, audio classification on MSPM0G5187, radar classification on IWRL6432, and F28E12 supports regression, forecasting and anomaly detection only.

• Prefer an NPU device (F28P55, MSPM0G5187, AM13E2) for larger convolutional models or tight latency budgets.

• Check the SDK example column if you want the quickest path to the board: those devices have ready-made SDK examples for all four time series tasks (see Deployment).

• Check memory early with estimate_memory.sh; the device’s Flash and RAM are on its ti.com product page.

• Existing designs: C2000 for motor control and power conversion, AM26x for industrial communication, CC devices for wireless sensors, MSPM0 for low-cost and low-power designs.