Ingenic (T33)

Ingenic T33 with Magik NPU

The Ingenic T33 SoC integrates a Magik NPU with ~0.6 TOPS theoretical performance. The soundSDK uses Magik's Uranus C API for NPU-accelerated inference.

Key Characteristics

Model Preparation Pipeline

# Step 1: Export ONNX model (Windows)
python export_model_onnx.py \
    --model path/to/model.h5 \
    --output cry_detect.onnx

# Step 2: Convert ONNX → .mgk (on Linux with Magik toolkit)
magik-transform-tools \
    --input cry_detect.onnx \
    --output cry_detect_model.mgk \
    --config magik_t33.cfg

Build

cd c/t33
./build.sh

The build links against:

• `liburanus_core.a` — NPU core inference engine

• `liburanus_tensor.a` — Tensor operations

• `liburanus_extend.a` — Extension operators

• `libaie.a` — AI Engine hardware driver

Deploy to Target

The Ingenic T33 board requires an HTTP server on the PC, then wget from the board:

# On PC: start HTTP server
python c/t33/http_server.py

# On T33 board: download files
wget http://192.168.0.xxx:8000/cry_detect -O /mnt/sda0/cry_detect
wget http://192.168.0.xxx:8000/cry_detect_model.mgk -O /mnt/sda0/cry_detect_model.mgk

# Run
cd /mnt/sda0
chmod +x cry_detect
./cry_detect -f test.wav -e magik -w cry_detect_model.mgk

SD Card Path

The T33 board mounts SD cards at /mnt/sda0/. Ensure your test files and models are placed there.