DX-AllSuite Architecture Overview

This guide provides a comprehensive overview of the DXNN (DEEPX Neural Network) SDK architecture, its core modules, and supported environments.

Why DX-AllSuite?

DX-AllSuite integrates the complex pipeline—from model optimization to target hardware deployment—into a Single Workflow.

• Zero-Code Deployment: Deploy logic verified on a desktop directly to edge devices without any code modifications.
• Resource Optimization: Reduce the time spent on environment setup and system integration, allowing you to focus on building high-performance AI applications.
• End-to-End Solution: Provides model compilation, simulation, runtime execution, and monitoring within a single package.

System Architecture & Core Modules

Figure. DXNN SDK Full Architecture Overview.

[1] AI Model Compile Environment (Host Platform)
This environment converts and optimizes trained AI models into DEEPX NPU-proprietary binaries (.dxnn). It supports the x86_64 PC environment.

• DX-COM (Compiler): The core compiler that generates .dxnn binaries based on ONNX models and user configuration (JSON) files. It uses proprietary algorithms to convert models into hardware-optimized instructions while minimizing accuracy loss, enabling low-latency and high-efficiency inference.
• DX-TRON (Visualizer): A GUI tool for visualizing and analyzing the structure of .dxnn models. It provides color-coded graphs to intuitively show workload distribution between the NPU and CPU, helping developers understand execution flow and optimize performance.
• DX-ModelZoo: A repository of pre-trained models optimized for DEEPX NPUs. By providing ONNX models, JSON configs, and pre-compiled .dxnn binaries, it allows developers to test and deploy models immediately without going through the compilation process.

[2] AI Model Runtime Environment (Target Platform)
This environment runs optimized .dxnn models on actual NPU hardware and integrates them into applications. It supports both x86_64 and aarch64 environments.

• DX-RT (Runtime): The core software that interacts with firmware and drivers to run .dxnn models on the DEEPX NPU. It manages model loading, I/O buffers, inference execution, and real-time monitoring, offering flexible control via C/C++ and Python APIs.
• DX-APP: Sample applications built on DX-RT. It provides reference code for key vision tasks (Object Detection, Face Recognition, Image Classification, etc.), enabling developers to quickly build unique AI apps using these templates.
• DX-Stream: A GStreamer-based custom plugin that connects real-time video data to DEEPX NPU inference. It provides a modular pipeline covering pre-processing, inference, and post-processing, optimized for high-performance vision AI applications like intelligent cameras.
• Driver & FW: Supports high-speed data communication via the PCIe interface and manages NPU resource scheduling and power. It ensures stability between hardware and software while maximizing the hardware's potential for a seamless inference environment.

Development Workflow
The path from a trained model to NPU acceleration follows four logical steps.

• Step 1. Model Preparation: Export a model trained in PyTorch or TensorFlow to the ONNX format.
• Step 2. Optimization (Host): Compile the model using DX-COM with quantization and hardware optimization to generate a .dxnn file.
• Step 3. Deployment (Target): Transfer the generated .dxnn file to the target device.
• Step 4. Execution (Target): Run high-speed inference on the DEEPX NPU via DX-RT.

Technical Compatibility (Support Matrix)

DX-AllSuite provides proven compatibility with the latest operating systems and a wide range of hardware architectures, ensuring a stable bridge between your development environment and edge deployment.

Hardware & OS (Platform)

The following table outlines the supported environments for both the compilation (Host) and execution (Target) phases.

Category	DX-Compiler (Host)	DX-Runtime (Target)
Architecture	x86_64	x86_64, aarch64
OS	Ubuntu 24.04/22.04/20.04, Fedora, Redhat, CentOS	Ubuntu 24.04/22.04/20.04/18.04, Debian 13/12, Windows 11/10
Languages	Python 3.8, 3.9, 3.10, 3.11, 3.12	Python 3.8 or higher, C++14 or higher (C++17 for MSVC/Windows)

Model & Software Ecosystem

Supported AI Frameworks
DX-AllSuite seamlessly integrates with industry-standard frameworks to minimize refactoring.

• Frameworks: Ultralytics (YOLO), TensorFlow, PyTorch, Keras
• Formats: ONNX (Primary exchange format)

Global AI Ecosystem Partners
We maintain strong technical partnerships to ensure the DXNN SDK operates flawlessly within various industry-leading platforms.

• Cloud & Platform: AWS (IoT Greengrass), Baidu, DeGirum
• Vision & Algorithms: Ultralytics (YOLO Series), CVEDIA
• VMS & Security: Network Optix (Nx), VCA (Applied Intelligence)
• Embedded OS: Wind River (VxWorks)

DEEPX ModelZoo & Supported Tasks

DEEPX ModelZoo is a comprehensive repository providing over 270 pre-validated models. It allows users to immediately verify performance and accuracy across various hardware profiles without the need for a manual compilation process.

Task	Representative Models
Image Classification	ResNet, ResNeXt, MobileNet, EfficientNet (Lite/V2), ViT/DeiT/BEiT, MobileViT, FastViT, CasViT, RegNet, ShuffleNet, VGG
Object Detection	YOLO families (YOLOv3–YOLOv11, YOLOX, YOLO26), SSD, EfficientDet, NanoDet, DamoYOLO
Instance Segmentation	YOLACT, YOLOv5-Seg, YOLOv8-Seg, YOLO26-Seg
Semantic Segmentation	DeepLabV3/DeepLabV3+, SegFormer, BiSeNet, UNet
Oriented Object Detection (OBB)	YOLO26-OBB
Zero-shot Instance Segmentation	FastSAM
Face Detection	RetinaFace, SCRFD, ULFGED, YOLOv5-Face, YOLOv7-Face
Face Recognition	ArcFace (IResNet50/100, MobileFaceNet, R50)
Face Landmark	TDDFA v2 (MobileNet variants)
Face Attribute	FaceAttrResNetV1-18
Pose Estimation (Human)	CenterPose, YOLO26-Pose, YOLOv8-Pose
Hand Landmark	MediaPipeHandsLite
Person Attribute	DeepMAR (ResNet18/50)
Depth Estimation	FastDepth, SCDepthV3
Image Denoising	DnCNN variants
Low-light Enhancement	ZeroDCE
Super-resolution	ESPCN (x2/x3/x4)

Key Features of the ModelZoo Pipeline

• YAML-Centric Orchestration: Manage pre-processing, post-processing, evaluation, and compilation settings in a single YAML file for high reproducibility and easy operation.
• Unified Workflow: Integrates list, info, eval, compile, and benchmark functions into a single CLI system, allowing flexible combination of stages.
• Optimized Multi-Profile Support: All models are quantized and optimized for the DEEPX NPU architecture, ensuring consistent validation from ONNX evaluation to NPU execution.
• Extensible Registry: Supports plugin-style extensions for pre/post-processing, datasets, and evaluators, enabling fast onboarding of custom models.
• Broad Task Coverage: Extensive support for service-ready CV tasks beyond basic classification and detection, including Face Analysis, OBB, and Image Enhancement.

---