2024高通AI白皮书-让AI触手可及-高通_78页_7mb

Qualcomm AI White Paper: Making AI Accessible through NPU and Heterogeneous Computing
The Qualcomm AI White Paper emphasizes the shift of AI computing from the cloud to end devices, driven by user demand, technical advancements, and improved user experience. Key points include:

1. Generative AI and End-Device Processing:

   • Generative AI (GAI) is transforming industries, requiring diverse processing capabilities. End devices like smartphones, PCs, and autonomous vehicles must handle complex GAI workloads, from real-time tasks (e.g., voice recognition) to continuous AI applications (e.g., predictive assistants).
   • Heterogeneous computing, combining CPUs, GPUs, and NPUs, is critical for optimizing performance, energy efficiency, and battery life. Qualcomm’s Hexagon NPU, designed for low-power AI inference, has evolved to support advanced models like StableDiffusion and Llama2.

2. Qualcomm’s NPU Innovation:

   • Hexagon NPU, first introduced in 2007, has seen continuous upgrades. By 2023, it achieved 98% performance improvement and 40% energy efficiency boost for GAI tasks.
   • Features like micro-tile inference, dedicated power supply, and large shared memory enhance AI capabilities. NPU’s ability to handle tensor operations (e.g., for GPT models) reduces memory bandwidth usage and improves scalability.

3. Hybrid AI: Bridging Cloud and Edge:

   • Hybrid AI distributes workloads between devices and the cloud, balancing cost, performance, and privacy. For example, small models run locally for low-latency tasks, while complex models leverage cloud resources.
   • This approach reduces cloud dependency, e.g., by enabling speculative decoding (partial computations in edge devices and corrections in the cloud), lowers costs, and maintains latency-sensitive operations.

4. AI Hardware and Software Ecosystem:

   • Qualcomm’s AI Engine integrates Hexagon NPU, Adreno GPU, and Kryo/Oryon CPU for seamless AI processing. The AI Software Stack (e.g., Qualcomm AI Studio) supports cross-platform development, with tools like AIMET for model optimization (quantization, pruning).
   • Snapdragon platforms (e.g., X Elite for PCs, 8 Gen 3 for mobiles) demonstrate leadership in AI performance, supporting GAI applications like 3D content creation and real-time super resolution.

5. Market Trends and Applications:

   • By 2027, 60% of PCs are expected to be AI-powered (AIPC). Smartphones will see rapid GAI adoption, with 55 billion units (43% of total) projected to support GAI by 2027.
   • Automotive applications include AI-driven assistants, autonomous driving (via Snapdragon Ride platform), and predictive maintenance. XR devices leverage Qualcomm’s AI for immersive 3D content and virtual avatars.

6. Responsible AI and Ecosystem Leadership:

   • Qualcomm emphasizes ethical AI, adhering to global regulations and privacy standards. Its AI solutions aim to enhance security and user trust.
   • With over 20 billion AI-enabled devices using Snapdragon and Qualcomm platforms, the company is positioned to scale hybrid AI globally, fostering ecosystem growth and innovation.

The paper underscores Qualcomm’s role in enabling AI accessibility by advancing hardware-software integration, optimizing NPU and heterogeneous architectures, and supporting cross-industry AI applications. It highlights the convergence of 5G, edge computing, and AI to drive a new era of intelligent devices and user-centric innovation.