稀疏计算

Search documents
AI算力集群迈进“万卡”时代 超节点为什么火了?
Di Yi Cai Jing· 2025-07-30 10:24
从千卡集群到万卡集群、十万卡集群,算力芯片集群越来越大,是因为大模型越来越大,参数量已经迈 向万亿。要怎么做好算力支撑,业内不得不思考这个问题。 超节点是一种应对大算力集群芯片协同问题的方案。通过整合算力芯片资源,在一个超节点内构建低延 迟、高带宽的算力实体,让单个超节点尽可能大地释放算力,避免芯片空闲,提高算力利用效率,支撑 千亿乃至万亿参数模型的训练和推理。 即便是在单颗算力芯片制程受限的情况下,超节点技术也能让集群性能继续提升。壁仞科技OCS超节点 项目相关负责人董朝锋告诉记者,在国内,做超节点已经是个明显的趋势,随着算力集群迈入"万卡协 同"时代,探讨未来算力的构建范式演进非常重要。 超节点为什么火了? 刚落幕的2025世界人工智能大会暨人工智能全球治理高级别会议(WAIC)上,华为展台展出一整面昇 腾384超节点的"墙"。记者看到,不断有观众与之合影。 超节点(SuperPod)成为本届WAIC的热点。除了华为,WAIC期间,上海仪电还联合曦智科技、壁仞科 技、中兴通讯发布国内首个光互连光交换GPU超节点光跃LightSphere X,新华三推出H3C UniPoD S80000超节点,超聚变也展出 ...
心智观察所:说芯片无需担忧,任正非战略思想有什么技术底气
Guan Cha Zhe Wang· 2025-06-10 07:02
Core Viewpoint - Huawei's founder Ren Zhengfei asserts that the company is not overly concerned about chip issues, claiming that through methods like "stacking and clustering," Huawei's computing capabilities can match global leaders in the field [1]. Group 1: Technological Innovations - The concept of "stacking and clustering" involves system-level innovations to compensate for the performance deficiencies of individual chips. Huawei's Ascend 910B chip exemplifies this approach, utilizing self-developed CCE communication protocols to create efficient clusters that support the training of large models, achieving computing power comparable to top GPUs [3]. - Huawei's algorithm optimization is notable, with the "using mathematics to supplement physics" philosophy leading to techniques like sparse computing and model quantization, which reduce hardware dependency. The MindSpore framework has lowered AI training computational demands by over 30% [4]. - The Chiplet technology reflects Huawei's strategic thinking in engineering practice, allowing the company to overcome generational gaps in single-chip processes through architectural innovation and system-level optimization [7]. Group 2: Competitive Strategies - Huawei's strategy mirrors AMD's rise, which focused on modular design and efficient interconnect technology rather than solely on process nodes. AMD's EPYC processors captured about 15% of the global server market in 2020, demonstrating the effectiveness of targeted optimizations in specific scenarios [5]. - The Chiplet architecture allows for the integration of multiple smaller chips manufactured with different process nodes, thus bypassing the limitations of single-chip advancements. This approach enables Huawei to achieve competitive performance and functionality without being constrained by the latest process technologies [8][9]. - Huawei's long-term investment in talent and education is a core strength, with approximately 114,000 R&D personnel and over 1.2 trillion yuan invested in R&D over the past decade. The "Genius Youth" program attracts top talent, ensuring a robust pipeline for innovation [9][10]. Group 3: Challenges and Future Outlook - Despite the advantages of cluster computing, challenges remain in energy consumption, costs, and communication bottlenecks. In scenarios requiring high single-thread performance, the benefits of clustering may not be fully realized [10]. - If Huawei continues to improve in chip manufacturing, supply chain stability, and global positioning, it could compete more effectively with international giants across a broader range of fields [10].