Software Stack Overview & Architecture
The A8550MA1 is a System-on-Module (SOM) based on the Qualcomm QCS8550. The module integrates processor, memory, and storage, can independently run Ubuntu 22.04, and connects to cameras, displays, networks, audio, and control peripherals via a companion base board or customer-designed carrier board.
In terms of AI processing capabilities, the typical usage is to perform data acquisition, preprocessing, AI inference, business processing, and result output locally on the A8550MA1. It can also provide AI services to customer host systems or business systems over the network or standard interfaces.
Software Stack Layers
| Layer | Main Components | Function |
|---|---|---|
| Scenario Application Layer | Machine vision, intelligent voice, multimodal interaction, RAG knowledge Q&A, robotics, and industrial applications | Implements interaction logic, business workflows, and user interfaces for final products |
| AI Services & Application Framework | AidVoice voice pipeline, AidGenSE LLM service, RAG service, AidStream media pipeline, customer-developed services | Combines model capabilities into callable, integrable scenario services |
| AI SDK & Acceleration Layer | AidLite, AidCV, AidStream, AidVoice, and other SDKs | Provides model inference, image processing, audio/video processing, and voice capability interfaces, invoking heterogeneous compute resources such as CPU, GPU, NPU/DSP |
| System & BSP Layer | APLUX Full Image, Ubuntu 22.04, device drivers, runtime libraries, containers, and development tools | Provides the system runtime environment, peripheral support, application deployment, and debugging capabilities |
| Hardware Platform Layer | QCS8550, LPDDR5x, UFS, and CPU, GPU, NPU/DSP, ISP, VPU | Handles general-purpose computing, AI inference, image signal processing, and audio/video codec |
Heterogeneous Computing Division
| Compute Unit | Primary Tasks | Typical Applications |
|---|---|---|
| CPU | Business logic, data scheduling, general-purpose computing, and service management | Application services, protocol handling, post-processing |
| GPU | Graphics rendering and image tasks suitable for parallel computing | UI rendering, image processing, some floating-point models |
| NPU/DSP | High-performance, low-power quantized model inference | Object detection, speech recognition, classification, on-device LLMs |
| ISP | Camera image signal processing | Multi-camera input, image quality processing |
| VPU | Hardware video encoding/decoding | Video analytics, recording, streaming |
Applications can select the appropriate compute unit based on model type, precision requirements, and real-time goals. SDKs like AidLite encapsulate the underlying inference backends, reducing the effort required to adapt applications directly to hardware interfaces.
Typical Deployment Modes
Integrated On-Device Deployment
Models, AI services, and business applications all run on the A8550MA1, directly connected to cameras, microphones, displays, and control peripherals. Suitable for smart terminals requiring low latency, offline operation, and data privacy.
On-Device AI Services
The A8550MA1 runs inference or LLM services locally and provides APIs to customer applications over interfaces such as Ethernet. The customer system handles the UI and business workflows, while the A8550MA1 handles compute-intensive AI tasks.
Customer Product Integration
The A8550MA1 is combined with a customer-designed carrier board, with sensors, cameras, displays, audio, and communication modules connected as needed by the product. Applications, models, and system images can be deployed and upgraded uniformly with the finished product.
From Model to Application
- Select or prepare a model: Obtain pre-adapted models from the Model Farm, or use the customer's own models.
- Convert and optimize: Use tools such as AIMO to complete model conversion, quantization, and platform adaptation.
- Deploy and infer: Load and run models on the A8550MA1 using components such as AidLite and AidGenSE.
- Scenario integration: Build complete applications by combining AidStream, AidVoice, RAG services, or customer business interfaces.
- Validation and delivery: Verify accuracy, latency, resource usage, stability, and exception recovery capabilities.
Software & Materials Delivery
| Category | Main Content |
|---|---|
| System Image | APLUX Full Image, including Ubuntu 22.04, base runtime environment, and companion components |
| BSP & Development Materials | Kernel, device tree, peripheral drivers, build, and flashing-related materials |
| AI Development Components | AidLite, AidCV, AidStream, AidVoice, AidGenSE, and other SDK or service components |
| Models & Tools | Model Farm, AIMO model optimization platform, and corresponding examples |
| Examples & Scenario Demos | Basic inference examples, LLM and RAG examples, intelligent voice, and other typical scenario demos |
Tip
Specific components, versions, and delivery formats are subject to the firmware packages, SDK packages, and release notes for the corresponding version.