“暴力计算”模式触及极限，算力进入系统工程时代

Core Insights - The computing power industry is undergoing a significant shift from a focus on single-point performance to system efficiency and multi-party collaboration in response to the demands of large models [1][2][3] Group 1: Industry Trends - The consensus among industry leaders is that the competition in computing power has evolved, necessitating a shift from a full-stack approach to a collaborative system engineering model [1][2] - As the scale of models increases to trillions of parameters, the challenges faced by computing systems extend beyond peak computing power to include interconnect bandwidth, storage hierarchy, power cooling, and system stability [2][3] - Traditional computing nodes are becoming inadequate for supporting large-scale models, leading to a consensus shift towards super-node and super-cluster models that utilize high-speed buses to connect multiple GPUs [3] Group 2: Challenges in the Ecosystem - The full-stack self-research model adopted by many domestic manufacturers has led to increased internal competition and fragmentation, creating multiple closed ecosystems that complicate user experiences [4][5] - Users face significant challenges in adapting to various chip architectures, leading to high costs and reduced development efficiency due to the need for extensive optimization and adaptation [5][6] - The lack of a cohesive ecosystem in domestic AI development is seen as a bottleneck, with manufacturers struggling to achieve seamless integration between hardware and software [6] Group 3: Shift to Open Computing - Open computing is being emphasized as a necessary approach, requiring manufacturers to move away from a "one company does it all" mentality towards a collaborative model where multiple firms contribute to different layers of the system [7][8] - The transition to open computing involves significant challenges, including the need to relinquish some control and profit margins, as well as establishing effective coordination mechanisms among various stakeholders [7][8] - A layered decoupling of the industry chain is essential for open computing, where different companies work on components like chips, interconnects, and storage while maintaining unified standards to ensure system efficiency [8] Group 4: Future Outlook - The coexistence of tightly coupled closed systems and open collaborative systems is expected to persist in the rich application landscape of the domestic market [9] - The ability to create an efficient, collaborative, and sustainably evolving system will be a critical factor determining the survival of manufacturers in the evolving landscape of large models and super clusters [9]