Introduction

Hardware

Genio Concept

The SoC used by NIO 12L is Genio 1200. Genio is the SoC series name used by MediaTek for IoT platforms.

Genio 1200 Architecture

Genio 1200 has the following hardware available for AI acceleration:

Hardware	Model	Description
GPU	Arm Mali-G57 MC5	Provides floating-point model acceleration
NPU	2xMDLA2.0 + 2xVP6	Neural Network Processing Unit

NPU is the collective name for MDLA and VPU. Inside the NPU, there are small, medium, and large-scale cores. This heterogeneous design provides support for various types of AI.

About MDLA

MDLA stands for Mediatek Deep Learning Accelerator. It is primarily used for convolutional neural network acceleration.

Limitations

MDLA2.0 does not support recurrent neural networks or Transformer-based generative large models.

Deployment

Overview

The Genio AI software stack supports both analytical AI and generative AI, using a host-device architecture. This means models are converted and quantized on the host, then the processed models are run on the device.

Two Modes

The inference modes running on the device are divided into online and offline modes:

Mode	Inference Framework	Model Format
online	General AI frameworks (TFlite, ONNX)	tflite, onnx
offline	Neuron Runtime	dla

The core difference is whether a general AI framework's inference parsing tool needs to be configured on the device. Both modes require configuring the NPU's Neuron Runtime, because when the general AI framework's inference parsing tool calls the NPU, it also goes through Neuron Runtime.

Online Mode

This mode is suitable for scenarios where models can efficiently utilize heterogeneous frameworks.

Analytical AI - Online Mode Hardware Scheduling

Framework	Supported Hardware
ONNX Runtime	CPU, NPU
TFlite Interpreter	CPU, GPU, NPU

For models in tflite format, MediaTek provides a dedicated conversion tool called NeuroPilot Converter, which is more friendly to hardware acceleration.

Generative AI

Limitations

Currently, only offline mode is available and is for educational purposes only. Genio 1200 does not support generative AI yet.

Generative AI Deployment Process

1. Quantize the model using the mtk_llm_sdk tool
2. Convert to tflite format using the post_training_quantize tool
3. Convert to dla format using NeuroPilot Converter
4. Compile into an executable and push to the device for execution
5. Run on the NPU via Neuron Runtime on the device

Software Stack

NeuroPilot

NeuroPilot (NP) is MediaTek's collection of tools, documentation, runtimes, and interfaces for AI development.

Version Information

For a given SoC, the NP version that can be used throughout its lifecycle is fixed, because the types of operators supported by the hardware were determined during the initial hardware design.

• MDLA2.0's "2.0" identifies its version number
• NP version needs to match the MDLA hardware version

Genio 1200 Support Status

Feature	Support Status
Analytical AI (online)	TFlite ✅ / ONNX ❌
Analytical AI (offline)	✅
Generative AI	❌