新华三集团交换机产品线产品管理部部长陈伯超解释，端网协同的核心是让 GPU 或网卡 "感知" 网络状态："我们通过与网卡对接来获取整网的通信 关系，并在网络侧针对GPU间的通信关系进行精准调优，实现高效协同效果。" 这一方案可适配多元芯片，无论何种交换机、GPU 或网卡，只要遵从方法论即可进行适配。目前，该方案已在 50 余款新华三交换机及多家 GPU、 网卡厂商的联合验证中成功运行 Llama3 模型，网络性能平均提升 16%，防拥塞效果提升 20%。 【环球网科技报道 记者 林梦雪】"在智算时代，网络在 AI 基础设施中的投资占比虽仅约 10%，但性能优化却能撬动超过 30%的算力业务能力。"新 华三集团高级副总裁、网络产品线总裁乔剡在第九届未来网络发展大会期间提到。 乔剡强调，这正是新华三从去年起将"算力×联接"作为核心战略的底层逻辑。"过去计算与网络生态相对割裂，而在智算时代，这种割裂状态必须被打 破。"他解释道，一个经过优化的网络能显著提升整体算力业务能力，其价值相当于"免费获取的算力增益"。 面对国产GPU生态多元化的行业现状，新华三秉持"多元异构"开放战略，与众多品牌厂家深度合作。乔剡表示，相比 ...

Core Insights - The Ninth Future Network Development Conference was held in Nanjing, focusing on global network technology and development, with participation from industry leaders and experts [1] - Unisplendour Corporation's subsidiary, H3C Group, showcased cutting-edge technologies in intelligent computing networks and emphasized its strategic approach to enhancing industrial intelligence through network and computing integration [1][2] Group 1: Intelligent Computing Network Innovations - H3C Group highlighted the need for proactive collaboration between computing and networking, moving away from traditional passive models to a more integrated approach [2][3] - The company introduced the concept of "end-network collaboration," which can potentially enhance overall computing capabilities by over 20% to 30% through optimized network adjustments [3][4] - H3C's solutions include advanced algorithms for intelligent network scheduling, significantly reducing the deployment time for large clusters from weeks to days [3][4] Group 2: DDC Architecture and Performance - H3C launched the DDC (Diversified Dynamic-Connectivity) architecture, which aims to decouple computing power from the network layer, allowing for efficient scheduling and integration of diverse computing resources [5][6] - The DDC architecture has been validated to improve effective bandwidth by up to 107% compared to traditional routing protocols, showcasing superior performance in demanding data transmission scenarios [5][6] - The introduction of the new 800G AI computing switch, H3C S9828-128EP, represents a significant advancement in network performance, energy efficiency, and operational cost reduction [6][7] Group 3: Commitment to Open Collaboration - H3C emphasizes an open and collaborative approach, allowing various hardware manufacturers to integrate into its ecosystem without stringent proprietary standards [4][6] - The company's strategy is to provide adaptable solutions that enhance performance across different hardware platforms, fostering a more inclusive computing environment [4][6] - H3C's ongoing commitment to innovation and practical solutions positions it as a leader in the digital transformation landscape, aiming to empower various industries through intelligent connectivity [7][8]

近日，以"网络全球，决胜未来"为主题的第九届未来网络发展大会在南京召开。这场由江苏省重点支持、南京市重点 打造的网络信息领域全球品牌盛会，汇聚了互联网研究任务组（IRTF）主席、多位两院院士、企业代表及顶尖专家， 共商未来网络技术与发展之道。 作为数字化及AI解决方案领导者，紫光股份旗下新华三集团深度参与大会，全面呈现在智算网络领域的前沿技术成 果。此外，新华三集团高级副总裁、网络产品线总裁乔剡在主论坛发表重磅演讲，清晰勾勒出在AI狂飙时代下，新华 三"以网强算"、助推产业智能化跃迁的技术路径与战略思考。 随着AI模型规模从万亿参数跃升至十万亿乃至数十万亿，远超单GPU算力的增长幅度，庞大的集群规模带来了更复杂 的流量动态与多品牌异构产品组成的多元算力接入需求。 新华三深刻洞察到，传统的、割裂的计算与网络生态已被打破，网络必须从被动承载转向主动协同。正是基于"算力× 联接"的核心理念，新华三在此次大会上重点展示了其通过端网协同与DDC（Diversified Dynamic-Connectivity，多元动 态联接）架构破解智算难题的创新实践，为构建敏捷、高效、安全的智能化基础设施提供了新华三方案。 端网协同 ...