Skip to main content

MediaTek NeuroPilot

Vendor: MediaTek

Category: NPU Neural Processor

Architecture: APU (AI Processing Unit)

Introduction

MediaTek NeuroPilot AI platform for AI acceleration in Dimensity and Kompanio series SoCs. Supports multi-framework fusion (TensorFlow/PyTorch/ONNX), providing heterogeneous computing capabilities.

Specifications

ModelComputeMemoryInterfaceTDPProcess
Dimensity 940040 TOPS (INT8)LPDDR5X (shared)Integrated SoC10W3nm
Dimensity 930032 TOPS (INT8)LPDDR5T (shared)Integrated SoC8W4nm

Official Website

Visit Official Website

Driver Downloads

Linux

OS Support

WindowsLinuxmacOSAndroid
✅ (Android)

Version History

VersionRelease DateDescription
NeuroPilot 9.02024On-device large model 7B inference

Performance Benchmarks

ModelTaskPerformance Metric
Dimensity 9400NPU TOPS~70 TOPS
Dimensity 9400Camera AIReal-time semantic segmentation
Dimensity 9300On-device LLM7B quantized inference

Pricing

ModelReference PriceNotes
Dimensity 9400Provided with SoCFlagship phone SoC

Quick Installation

Android

MediaTek NeuroPilot is automatically invoked through Android NNAPI. TFLite models directly benefit from NPU acceleration.

# Use NeuroPilot SDK for model optimization
# Obtain from MediaTek Developer Portal

Code Examples

Python (TFLite Android)

import tflite_runtime.interpreter as tflite

# Android: Use Dimensity NPU via NNAPI
interpreter = tflite.Interpreter(
model_path="model.tflite",
experimental_delegates=[tflite.load_delegate('libneuropilot_delegate.so')]
)

Architecture Highlights

  • APU (AI Processing Unit): MediaTek integrated NPU, co-processing AI tasks with CPU/GPU
  • NeuroPilot SDK: Developer toolchain supporting model optimization and deployment
  • On-Device LLM: Dimensity 9400 supports 7B parameter LLM on-device inference

Model Compatibility

Model/FrameworkSupportNotes
NNAPI✅ NativeAndroid standard
TFLiteNeuroPilot optimized
On-Device LLM7B quantization
Camera AIBest for camera scenarios

If you're evaluating alternatives, the following products may also fit your scenario: