容芯致远石旭：智算时代呼唤以GPU为核心的AI体系结构

Core Insights - The demand for computing power has surged since the advent of generative AI in 2023, with high-end GPUs becoming essential production materials as AI applications proliferate by 2025 [1] - The transition from traditional computing to intelligent computing is marked by the introduction of the Deepseek mixture of experts (MoE) model, which allows for on-demand activation of sub-models, reducing training and inference costs [2] - The industry is moving towards a new data center architecture that utilizes electrical computing and optical transmission to overcome the limitations of traditional copper interconnects [3] Industry Trends - In 2023, China sold 4.7 million traditional computing servers and 150,000 intelligent computing servers, with projections indicating a decline in traditional server sales to over 1 million and an increase in intelligent server sales to over 100,000 by mid-2025 [1] - The shift towards GPU-centric architectures is driven by the limitations of CPUs in handling large-scale parallel computing tasks, with GPUs becoming the core of AI computing systems [4][5] - The industry is innovating to address the high failure rates and energy consumption of GPUs, with advancements in GPU hot-swapping and RAID technologies aimed at enhancing reliability and extending lifespan [5] Technological Innovations - The AGC (AI computer system with GPU as its Core) architecture proposed by companies like容芯致远 positions the CPU as a peripheral, reducing reliance on high-performance CPUs and allowing for better performance from domestic GPUs [5] - The effective utilization of computing power (MFU) can be improved from an average of 40% in traditional servers to over 60% through architectural innovations, enhancing overall computational efficiency [6]