CPU，为何“偷偷转型”？

Core Insights - The Yole Group's report indicates a significant milestone where GPU sales are projected to surpass CPU sales for the first time in 2024, marking a new era in the semiconductor industry dominated by accelerated computing [1] - The shift in computational focus towards GPUs, NPUs, and ASICs raises questions about the future role of traditional CPUs in large-scale parallel computing tasks [1] - The demand for CPUs is evolving, as they transition from simple logic controllers to central scheduling units in heterogeneous systems, impacting market dynamics and capital flows from data centers to edge devices [1] Group 1: CPU Challenges and Transformation - Traditional CPU-centric architectures face efficiency issues in managing data processing workflows, particularly under AI workloads, leading to increased system costs and power consumption [2] - The reliance on speculative execution in modern CPUs presents challenges when handling AI and machine learning tasks, resulting in wasted energy and delays due to frequent pipeline flushes [2] Group 2: Innovations in Processor Architecture - The industry is witnessing a shift towards de-speculative microarchitecture, exemplified by a newly patented deterministic execution model that enhances efficiency in matrix computations while maintaining compatibility with standard instruction sets [3] - System-level architecture is evolving with the introduction of Network Attached Processing Units (NAPUs) to alleviate I/O bottlenecks by offloading specific tasks from the CPU to dedicated hardware [3] Group 3: Market Dynamics and CPU Applications - Despite the rising demand for GPUs in training, the inference market is becoming increasingly sensitive to cost and efficiency, creating opportunities for new CPU designs [5][6] - The U.S. data center CPU demand is expected to grow at a compound annual growth rate (CAGR) of 7.4%, driven by the economic realities of AI application deployment [6] - CPUs are becoming essential for AI inference tasks, especially for mid-sized models, as they can leverage underutilized resources in public cloud environments, offering significant total cost of ownership (TCO) advantages [6] Group 4: Evolving Role of CPUs in AI - The demand for memory capacity driven by large AI models is reshaping the market value of CPUs, as they increasingly serve as L4 caches for GPUs, enhancing overall system performance [7] - In edge computing and smart devices, the need for heterogeneous collaboration is surpassing single-chip performance, with CPUs handling low-latency tasks while GPUs and NPUs manage high-concurrency computations [7][8] Group 5: Competitive Landscape in the Processor Industry - The processor industry is experiencing a competitive reshaping, with startups focusing on AI-specific architectures emerging alongside traditional giants adapting their strategies [9] - NeuReality's NR1 chip exemplifies the trend towards specialized architectures, aiming to address traditional CPU bottlenecks in AI data processing and significantly improving TCO [9] - Major players like NVIDIA are investing heavily in x86 ecosystems, indicating the continued strategic importance of high-performance x86 CPUs in heterogeneous computing environments [10] Group 6: Future of CPU Architectures - The Arm architecture is gaining traction in the server market, projected to capture 21.1% of global server shipments by 2025, driven by cloud providers' in-house chip developments [11] - The coexistence of x86 and Arm architectures, along with the integration of general-purpose and specialized AI CPUs, is defining a complex ecosystem where competitive advantage will depend on architectural openness and efficiency in heterogeneous computing [11]