我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
从重装旅到弹性军团AI算力集群化的必然之路
视频
音频
原始脚本
从重装旅到弹性军团,AI 算力集群化的必然之路。 当 NVIDIA 将72颗 GB200 GPU 塞进单个机柜,打造出 NVL72系统时,看似抵达了单节点算力巅峰,实则暴露了集中式架构的局限。 与此同时,华为 Cloud Matrix 384超节点以16机柜全光互联形态落地,MD、NVIDIA 联合 Meta 等成立,ESUN 联盟推动以太网开放化。 三条技术路线的交织印证着一个结论,AI 算力正告别大马拉小车的集中式时代,迈向集群化、弹性化的新征程。 这一转变是算力需求演进、成本逻辑重构与历史规律重演共同作用的必然结果。 一、集中式算力的重装旅困境、性能顶点与效率悖论。 NVIDIA GB200 NVL 72的单机柜72 GPU 架构,将集中式算力的优势与短板展现的淋漓尽致。 这款 AI 时代大型机凭借 NVLink 同缆互联实现亚微秒级延迟,单柜算力达176 FLOPS,在稠密大模型训练中表现强悍。 但正如上世纪70年代的大型机,这种重装铝式设计正逼近物理与经济的双重极限。 从物理约束看,72颗 GPU 的密集部署已触及工程天花板。 单柜120千瓦的功耗是传统 H100服务器的10倍。 即便采用直接芯片液冷,散热效率提升空间也有限。 更关键的是互联瓶颈,铜缆 NVLink 传输距离限制在2米以内,一旦超出单机柜范围,通信效率便断崖式下跌。 系统扩展能力被锁死在机柜边界内。 经济与效率层面的悖论更为突出。 GB 200 NVL 72单套成本高达300万美元,却面临严重资源浪费风险。 就像载重百吨的卡车难以适配多数中小型货物,72 GPU 的密集算力面对大量轻量化推理任务时,必然陷入大马拉小车的尴尬。 某金融机构实测显示,处理日常风控模型推理时,其算力利用率不足20%。 更致命的是专有协议税,选择 NVLink 架构就必须配套 NVIDIA 的交换机、软件站等全链条产品,长期扩展成本成指数级增长。 这种困境是集中式架构的宿命,当单节点算力密度达到物理极限,其灵活性缺失,成本高企的弊端便成为 为不可逾越的障碍。 正如大型机最终被 PC 集群取代的历史轨迹,AI 算力也必然走向分布式重构。 二,集群化转向的底层逻辑,需求进化与历史的重演。 算力集群化并非突发的技术革命,而是需求驱动下的历史必然。 从华为 Cloud Matrix 384的落地到 ESUN 联盟的成立,背后折射出三重不可逆转的趋势。 需求端的分化催生了对弹性算力的迫切需求。 大模型时代的算力需求呈现双峰分布,一端是需要千卡级协同的 MoE 稀疏模型训练,另一端是海量碎片化的推理任务。 华为 Cloud Matrix 384的 MoE 亲和架构通过384卡全对等互联实现一卡一专 专家的分布式推理效率相比传统模式提升数倍。 E SUN 联盟推动的开放以太网架构更让动态组合算力成为可能。 Meta 演示的 DSF 架构已能无阻塞连接18432个 XPU,可根据任务大小灵活调整集群规模。 这种弹性军团模式完美解决了集中式架构无法适配多样化需求的痛点。 技术突破让集群化具备了工程可行性。 华为通过 Matrix Link 技术构建的双层网络架构,破解了大规模集群的通信瓶颈。 超节点内部二8T 卡间带宽实现纳秒级时延,跨节点网络支持微秒级互联,300 FLOPS 的算力规模比业界同类产品领先67%。 更关键的是光模块技术的迭代,为机柜间光互联的普及扫清了障碍。 这些技术创新让集群化从理论可行变为工程可靠,历史规律的重演给出了最明确的方向。 计算机产业已两次验证分布式战胜集中式的逻辑。 上世纪80年代,PC 集群取代大型机。 2010年代,云服务器集群取代小型机。 如今,AI 算力的眼镜正复制这一路径。 华为在芜湖、贵安等地部署的超节点集群,本质上是算力机构 对话、机柜集群化的现代实践。 E SUN 联盟用开放以太网连接易购 XPU 的尝试,与当年 TCP/IP 协议统一互联网的逻辑如出一辙。 技术路线可能迭代,但分散部署集中调度的效率优势永远成立。 三、技术路线的殊途同归,开放与混合架构的未来。 当前算力领域的三大技术路线看似分野,实则都在向开放化加混合互联的终极形态收敛。 这种收敛不是简单的技术统一,而是对性能、成本、灵活性三角难题的最优解探索。 华为 Cloud Matrix 384的架构创新已展现出混合互联的前瞻性。 其全光互联架构并非一味追求长距离,而是针对数据中心场景优化。 超节点内部通过 Scale Up 总线实现384卡无阻塞互联,跨节点通过 Scale Up 网络弹性扩展,既保留了光通信高带宽的优势,又通过分层设计控制了成本。 华为的开发放姿态,王 网络设备严格遵循国际标准,支持与第三方设备兼容对接,与 ESUN 联盟的理念不谋而合。 若未来华为将 Matrix Link 协议进一步开放,并融入 ASEN 的以太网优化方案,有望形成覆盖机柜内、跨机柜、跨数据中心的全场景互联体系。 E S U N 联盟的核心贡献在于打破了专有协议的垄断。 过去 NVIDIA 的 NVLink、Mellanox 的 InfiniBand 形成了封闭生态,企业一旦入局便难以脱身。 E S U N 通过优化 L 二、L 三以太网的帧结构与交换机制,让标准以太网具备了亚微秒延迟、无损传输的能力,首次在纵向扩展场景对专有协议形成挑战。 这种开放化带来的直接收益是成本降低,采用 ESUN 方案的集群硬件采购成本可降 降低30%以上,且能兼容 AMD、Intel 等不同厂商的 XPU。 这恰好弥补了华为架构在短距离互联成本上的短板,为混合架构提供了底层支撑。 Nvidia 的战略转向更凸显了趋势的必然性。 作为专有协议的既得利益者,其加入 ESUN 联盟的举动实则是对集中式架构局限的妥协。 未来 NVIDIA 大概率会形成单机柜 NVLink 保性能、跨机柜以太网求扩展的混合策略。 这与华为短距优化互联、加长距光互联的思路殊途同归。 三种路线的收敛指向同一个未来。 以开放以太网为基础协议,在机柜内采用低成本优化互联技术,跨机柜及数据中心采用光通信,通过软件定义实现算力的动态调度与弹性组合。 华为的 EMS 弹性内存存储技术已展现出这种潜力,通过内存池化实现算力与显存解绑,配合潮退夜训的调度策略。 可将资源利用率提升30%以上,这种硬件池化加软件调度的模式,正是集群化的核心价值所在。 四、结语。 算力民主化的必经之路,从 GB200的重装旅,到 CloudMatrix 384的弹性军团,再到 ESUN 联盟的开放生态,AI 算力的发展逻辑正在发生根本性转变。 衡量算力实力的标准不再是单节点的峰值性能,而是大规模集群的扩展能力。 资源利用率与成本可控性。 这一转变的本质是算力从贵族专属走向普惠公用的民主化进程。 华为的集群化实践证明了工程落地的可行性, E SUN 的开放联盟降低了行业准入门槛, NVIDIA 的转向则确认了趋势的必然性。 三者共同推动的集群化浪潮不仅将破解大模型训练的算力瓶颈,更将让 AI 算力像水电一样随取随用。 正如华为在全国三大枢纽数据中心实现的即开即用模式,10毫秒时延圈覆盖19个城市群,让中小企业也能享受到超大规模算力服务。 历史不会简单重复,但总会压着相同的韵脚。 上世纪末,PC 集群打破了大型机的垄断,催生了互联网革命。 今天,AI 算力集群正在打破专有架构的壁垒,为通用人工智能的落地铺平道路。 这条集群化之路既是技术演进的必然,更是算力民主化的必经之路。
修正脚本
从重装旅到弹性军团,AI 算力集群化的必然之路。 当 NVIDIA 将72颗 GB200 GPU 塞进单个机柜,打造出 NVL72系统时,看似抵达了单节点算力巅峰,实则暴露了集中式架构的局限。 与此同时,华为 Cloud Matrix 384超节点以16机柜全光互联形态落地,AMD、NVIDIA 联合 Meta 等成立ESUN 联盟,推动以太网开放化。 三条技术路线的交织印证着一个结论,AI 算力正告别大马拉小车的集中式时代,迈向集群化、弹性化的新征程。 这一转变是算力需求演进、成本逻辑重构与历史规律重演共同作用的必然结果。 一、集中式算力的重装旅困境、性能顶点与效率悖论。 NVIDIA GB200 NVL 72的单机柜72 GPU 架构,将集中式算力的优势与短板展现得淋漓尽致。 这款 AI 时代大型机凭借 NVLink 同缆互联实现亚微秒级延迟,单柜算力达176 FLOPS,在稠密大模型训练中表现强悍。 但正如上世纪70年代的大型机,这种重装旅式设计正逼近物理与经济的双重极限。 从物理约束看,72颗 GPU 的密集部署已触及工程天花板。 单柜120千瓦的功耗是传统 H100服务器的10倍。 即便采用直接芯片液冷,散热效率提升空间也有限。 更关键的是互联瓶颈,铜缆 NVLink 传输距离限制在2米以内,一旦超出单机柜范围,通信效率便断崖式下跌。 系统扩展能力被锁死在机柜边界内。 经济与效率层面的悖论更为突出。 GB 200 NVL 72单套成本高达300万美元,却面临严重资源浪费风险。 就像载重百吨的卡车难以适配多数中小型货物,72 GPU 的密集算力面对大量轻量化推理任务时,必然陷入大马拉小车的尴尬。 某金融机构实测显示,处理日常风控模型推理时,其算力利用率不足20%。 更致命的是专有协议税,选择 NVLink 架构就必须配套 NVIDIA 的交换机、软件栈等全链条产品,长期扩展成本成指数级增长。 这种困境是集中式架构的宿命,当单节点算力密度达到物理极限,其灵活性缺失、成本高企的弊端便成为不可逾越的障碍。 正如大型机最终被 PC 集群取代的历史轨迹,AI 算力也必然走向分布式重构。 二、集群化转向的底层逻辑,需求进化与历史的重演。 算力集群化并非突发的技术革命,而是需求驱动下的历史必然。 从华为 Cloud Matrix 384的落地到 ESUN 联盟的成立,背后折射出三重不可逆转的趋势。 需求端的分化催生了对弹性算力的迫切需求。 大模型时代的算力需求呈现双峰分布,一端是需要千卡级协同的 MoE 稀疏模型训练,另一端是海量碎片化的推理任务。 华为 Cloud Matrix 384的 MoE 亲和架构通过384卡全对等互联实现一卡一专家,分布式推理效率相比传统模式提升数倍。 E SUN 联盟推动的开放以太网架构更让动态组合算力成为可能。 Meta 演示的 DSF 架构已能无阻塞连接18432个 XPU,可根据任务大小灵活调整集群规模。 这种弹性军团模式完美解决了集中式架构无法适配多样化需求的痛点。 技术突破让集群化具备了工程可行性。 华为通过 Matrix Link 技术构建的双层网络架构,破解了大规模集群的通信瓶颈。 超节点内部达8T 卡间带宽实现纳秒级时延,跨节点网络支持微秒级互联,300 FLOPS 的算力规模比业界同类产品领先67%。 更关键的是光模块技术的迭代,为机柜间光互联的普及扫清了障碍。 这些技术创新让集群化从理论可行变为工程可靠,历史规律的重演给出了最明确的方向。 计算机产业已两次验证分布式战胜集中式的逻辑。 上世纪80年代,PC 集群取代大型机。 2010年代,云服务器集群取代小型机。 如今,AI 算力的演进正复制这一路径。 华为在芜湖、贵安等地部署的超节点集群,本质上是算力架构从单机到机柜集群化的现代实践。 E SUN 联盟用开放以太网连接异构XPU 的尝试,与当年 TCP/IP 协议统一互联网的逻辑如出一辙。 技术路线可能迭代,但分散部署集中调度的效率优势永远成立。 三、技术路线的殊途同归,开放与混合架构的未来。 当前算力领域的三大技术路线看似分野,实则都在向开放化加混合互联的终极形态收敛。 这种收敛不是简单的技术统一,而是对性能、成本、灵活性三角难题的最优解探索。 华为 Cloud Matrix 384的架构创新已展现出混合互联的前瞻性。 其全光互联架构并非一味追求长距离,而是针对数据中心场景优化。 超节点内部通过 Scale Up 总线实现384卡无阻塞互联,跨节点通过 Scale Up 网络弹性扩展,既保留了光通信高带宽的优势,又通过分层设计控制了成本。 华为的开放姿态,对网络设备严格遵循国际标准,支持与第三方设备兼容对接,与 ESUN 联盟的理念不谋而合。 若未来华为将 Matrix Link 协议进一步开放,并融入 ESUN 的以太网优化方案,有望形成覆盖机柜内、跨机柜、跨数据中心的全场景互联体系。 E S U N 联盟的核心贡献在于打破了专有协议的垄断。 过去 NVIDIA 的 NVLink、Mellanox 的 InfiniBand 形成了封闭生态,企业一旦入局便难以脱身。 E S U N 通过优化 L2、L3以太网的帧结构与交换机制,让标准以太网具备了亚微秒延迟、无损传输的能力,首次在纵向扩展场景对专有协议形成挑战。 这种开放化带来的直接收益是成本降低,采用 ESUN 方案的集群硬件采购成本可降低30%以上,且能兼容 AMD、Intel 等不同厂商的 XPU。 这恰好弥补了华为架构在短距离互联成本上的短板,为混合架构提供了底层支撑。 Nvidia 的战略转向更凸显了趋势的必然性。 作为专有协议的既得利益者,其加入 ESUN 联盟的举动实则是对集中式架构局限的妥协。 未来 NVIDIA 大概率会形成单机柜 NVLink 保性能、跨机柜以太网求扩展的混合策略。 这与华为短距优化互联、加长距光互联的思路殊途同归。 三种路线的收敛指向同一个未来。 以开放以太网为基础协议,在机柜内采用低成本优化互联技术,跨机柜及数据中心采用光通信,通过软件定义实现算力的动态调度与弹性组合。 华为的 EMS 弹性内存存储技术已展现出这种潜力,通过内存池化实现算力与显存解绑,配合错峰夜训的调度策略。 可将资源利用率提升30%以上,这种硬件池化加软件调度的模式,正是集群化的核心价值所在。 四、结语。 算力民主化的必经之路,从 GB200的重装旅,到 CloudMatrix 384的弹性军团,再到 ESUN 联盟的开放生态,AI 算力的发展逻辑正在发生根本性转变。 衡量算力实力的标准不再是单节点的峰值性能,而是大规模集群的扩展能力、资源利用率与成本可控性。 这一转变的本质是算力从贵族专属走向普惠公用的民主化进程。 华为的集群化实践证明了工程落地的可行性, E SUN 的开放联盟降低了行业准入门槛, NVIDIA 的转向则确认了趋势的必然性。 三者共同推动的集群化浪潮不仅将破解大模型训练的算力瓶颈,更将让 AI 算力像水电一样随取随用。 正如华为在全国三大枢纽数据中心实现的即开即用模式,10毫秒时延圈覆盖19个城市群,让中小企业也能享受到超大规模算力服务。 历史不会简单重复,但总会压着相同的韵脚。 上世纪末,PC 集群打破了大型机的垄断,催生了互联网革命。 今天,AI 算力集群正在打破专有架构的壁垒,为通用人工智能的落地铺平道路。 这条集群化之路既是技术演进的必然,更是算力民主化的必经之路。
英文翻译
From the Heavy Brigade to the Elastic Legion: The Inevitable Path of AI Computing Clusterization. When NVIDIA crammed 72 GB200 GPUs into a single cabinet to create the NVL72 system, it seemed to have reached the pinnacle of single-node computing power, but in reality, it exposed the limitations of centralized architecture. At the same time, Huawei’s Cloud Matrix 384 supernode was deployed in a 16-cabinet all-optical interconnect configuration, while AMD, NVIDIA, and Meta jointly established the ESUN Alliance to promote open Ethernet. The intersection of these three technical routes confirms a conclusion: AI computing power is bidding farewell to the centralized era of "a large horse pulling a small cart" and moving towards a new journey of clusterization and elasticity. This transformation is the inevitable result of the evolution of computing power demand, the reconstruction of cost logic, and the recurrence of historical patterns. **I. The Heavy Brigade Dilemma of Centralized Computing: Performance Peaks and Efficiency Paradoxes** NVIDIA's GB200 NVL 72 single-cabinet 72-GPU architecture fully demonstrates the advantages and drawbacks of centralized computing power. This AI-era mainframe, with sub-microsecond latency via NVLink copper cable interconnection and a single-cabinet computing power of 176 FLOPS, performs powerfully in dense large-model training. But just like the mainframes of the 1970s, this heavy brigade design is approaching both physical and economic limits. From a physical constraints perspective, the dense deployment of 72 GPUs has hit engineering ceilings. The single-cabinet power consumption of 120 kilowatts is ten times that of a traditional H100 server. Even with direct chip liquid cooling, the potential for improving heat dissipation efficiency is limited. More critically, the interconnection bottleneck: copper NVLink transmission distances are limited to within 2 meters; once beyond the single cabinet, communication efficiency drops off a cliff. System scalability is locked within cabinet boundaries. The paradox at the economic and efficiency levels is even more prominent. A single set of GB200 NVL72 costs up to $3 million but faces serious risks of resource waste. Just as a truck with a hundred-ton capacity is unsuitable for most small and medium-sized cargo, the dense computing power of 72 GPUs inevitably falls into the awkward situation of "a large horse pulling a small cart" when handling a large number of lightweight inference tasks. Real-world tests at a financial institution showed that when processing daily risk model inference, its computing power utilization was less than 20%. More deadly is the proprietary protocol tax: choosing the NVLink architecture requires the full chain of NVIDIA switches, software stacks, etc., and long-term expansion costs grow exponentially. This dilemma is the fate of centralized architecture. When the computing density of a single node reaches its physical limit, the drawbacks of inflexibility and high costs become insurmountable obstacles. Just as mainframes were eventually replaced by PC clusters, AI computing power is inevitably heading towards distributed reconstruction. **II. The Underlying Logic of Clusterization: Demand Evolution and Historical Repetition** Computing clusterization is not a sudden technological revolution but a historical inevitability driven by demand. From the deployment of Huawei's Cloud Matrix 384 to the establishment of the ESUN Alliance, three irreversible trends are reflected. Differentiation in demand has created an urgent need for elastic computing power. In the era of large models, computing power demand shows a bimodal distribution: on one end is MoE sparse model training requiring thousands of cards working in coordination; on the other end is massive fragmented inference tasks. Huawei Cloud Matrix 384's MoE-friendly architecture, with 384 cards fully interconnected in a peer-to-peer manner achieving "one card, one expert," improves distributed inference efficiency several times over traditional models. The open Ethernet architecture promoted by the ESUN Alliance makes dynamic combination of computing power possible. Meta's demonstrated DSF architecture can already connect 18,432 XPUs without blocking, flexibly adjusting cluster scale according to task size. This elastic legion model perfectly solves the pain point of centralized architecture's inability to meet diverse demands. Technological breakthroughs have made clusterization engineeringly feasible. Huawei's dual-layer network architecture built with Matrix Link technology has cracked the communication bottleneck of large-scale clusters. Inside a supernode, inter-card bandwidth of 8T achieves nanosecond-level latency; cross-node networks support microsecond-level interconnection, with a computing power scale of 300 FLOPS, leading industry peers by 67%. More critically, the iteration of optical module technology has cleared the obstacles for widespread adoption of inter-cabinet optical interconnects. These technological innovations have turned clusterization from theoretically feasible to engineeringly reliable. Historical repetition provides the clearest direction. The computer industry has twice verified the logic of distributed systems defeating centralized ones: in the 1980s, PC clusters replaced mainframes; in the 2010s, cloud server clusters replaced minicomputers. Today, the evolution of AI computing power is copying this path. Huawei's supernode clusters deployed in Wuhu, Guian, etc., are essentially the modern practice of transitioning from single machines to cabinet-level clusters. The ESUN Alliance's attempt to connect heterogeneous XPUs with open Ethernet is exactly the same logic as the TCP/IP protocol unifying the internet. Technical routes may iterate, but the efficiency advantage of decentralized deployment with centralized scheduling always holds. **III. Different Technical Routes Converging: The Future of Open and Hybrid Architectures** The three major technical routes in the current computing power field seem divergent, but in reality, they are all converging towards the ultimate form of openness plus hybrid interconnection. This convergence is not a simple technical unification but an exploration of the optimal solution to the triangle problem of performance, cost, and flexibility. Huawei's Cloud Matrix 384 architectural innovation has already shown the foresight of hybrid interconnection. Its all-optical interconnect architecture does not blindly pursue long distances but optimizes for data center scenarios. Inside the supernode, 384 cards are interconnected without blocking via the Scale Up bus; cross-node elastic expansion is achieved through the Scale Out network, retaining the advantages of high optical communication bandwidth while controlling costs through layered design. Huawei's openness—strictly following international standards for network equipment and supporting compatibility with third-party devices—coincides with the ESUN Alliance's philosophy. If Huawei further opens the Matrix Link protocol and integrates it into ESUN's Ethernet optimization solutions, it may form a full-scenario interconnection system covering intra-cabinet, inter-cabinet, and cross-data-center scenarios. The core contribution of the ESUN Alliance lies in breaking the monopoly of proprietary protocols. In the past, NVIDIA's NVLink and Mellanox's InfiniBand formed closed ecosystems; once enterprises entered, they found it hard to leave. By optimizing the frame structure and switching mechanism of L2/L3 Ethernet, ESUN has given standard Ethernet sub-microsecond latency and lossless transmission capabilities, for the first time challenging proprietary protocols in scale-up scenarios. The direct benefit of this openness is cost reduction: adopting ESUN solutions can reduce cluster hardware procurement costs by over 30% while supporting XPUs from different vendors like AMD and Intel. This precisely compensates for the short-distance interconnect cost disadvantage of Huawei's architecture, providing underlying support for hybrid architectures. NVIDIA's strategic shift further highlights the inevitability of the trend. As a beneficiary of proprietary protocols, its joining of the ESUN Alliance is actually a compromise on the limitations of centralized architecture. In the future, NVIDIA will likely adopt a hybrid strategy: NVLink within a single cabinet for performance, and Ethernet for inter-cabinet expansion. This converges with Huawei's approach of short-distance optimized interconnection plus long-distance optical interconnection. The convergence of the three routes points to one future: using open Ethernet as the base protocol, adopting low-cost optimized interconnection within cabinets, using optical communication across cabinets and data centers, and achieving dynamic scheduling and elastic combination of computing power through software definition. Huawei's EMS (Elastic Memory Storage) technology has shown this potential: by pooling memory to decouple computing power from video memory, combined with off-peak night training scheduling strategies, resource utilization can be improved by over 30%. This model of hardware pooling plus software scheduling is the core value of clusterization. **IV. Conclusion: The Necessary Path to Computing Democratization** From GB200's Heavy Brigade to Cloud Matrix 384's Elastic Legion, and then to the ESUN Alliance's open ecosystem, the development logic of AI computing power is undergoing a fundamental shift. The standard for measuring computing power strength is no longer the peak performance of a single node but the scalability, resource utilization, and cost controllability of large-scale clusters. The essence of this shift is the democratization of computing power—from aristocratic exclusivity to universal public utility. Huawei's clusterization practice proves the feasibility of engineering implementation; the ESUN open alliance lowers industry barriers; NVIDIA's turn confirms the trend's inevitability. The wave of clusterization jointly driven by these three forces will not only break the computing bottleneck of large-model training but also make AI computing power as readily available as water and electricity. Just as Huawei's "ready-to-use" model in its three major hub data centers, with a 10-millisecond latency circle covering 19 city clusters, allows small and medium-sized enterprises to enjoy ultra-large-scale computing services. History does not simply repeat itself, but it often rhymes. At the end of the last century, PC clusters broke the monopoly of mainframes, sparking the internet revolution. Today, AI computing clusters are breaking the barriers of proprietary architectures, paving the way for the realization of general artificial intelligence. This path of clusterization is not only inevitable in technological evolution but also the necessary road to computing democratization.
back to top