AMD Pensando™ Vulcano 800 AI NIC:专为横向扩展与跨域扩展而打造
Jul 23, 2026
不断变化的工作负载需求
随着 AI 模型的规模和复杂性不断增长,网络在决定集群能否发挥理论计算能力方面扮演着日益重要的角色。多节点训练和分布式推理部署如今已成为常态,但目前支撑大多数集群的网络架构并非为这一现实而设计。因此,GPU 在等待数据时会产生大量计算空闲时间,直接导致运营成本上升和迭代周期延长。要解决这一问题,需要重新思考集群的连接方式,并将 AI NIC 作为基础设施决策中的首要考量。
AMD Pensando™ Vulcano 800 AI NIC 专为这种规模的集群而打造。每个 AI NIC 可提供 800 Gbps 连接能力,并支持独特的每 GPU 配置 3 个 NIC,从而为每个 GPU 提供高达 2.4 Tbps 的横向扩展带宽,为训练和分布式推理工作负载提供所需的吞吐余量。该 AI NIC 并未将高带宽视为单一维度的指标,而是围绕两种共同决定真实集群效率的不同网络架构进行设计:横向扩展(scale-out)和跨域扩展(scale-across)。
横向扩展与跨域扩展
横向扩展和跨域扩展描述了两种不同但互补的连接需求,两者共同决定集群处理流行的 AI 工作负载的能力。
横向扩展是指将工作负载分布到多个节点的能力,使有效 GPU 资源池超越单台服务器所能提供的规模。对于大模型训练,这意味着要在数十乃至数百个节点之间协调梯度同步,同时将停顿时间降至最低。对于推理,这意味着要在一组服务实例之间分发请求,同时避免网络成为限制因素。在这一层面,节点间带宽和延迟直接决定新增硬件能否带来成比例的吞吐量提升,抑或随着通信开销增加而出现收益递减。
跨域扩展则将连接范围延伸到机架之外。随着 AI 部署规模扩大,集群越来越多地跨越多个数据中心。在此类架构中,数据中心之间承载的通信与单个设施内机架内部的通信同样关乎性能。因此,站点之间的延迟和带宽会带来与站点内部相同的影响:吞吐量下降、作业完成时间延长,以及 GPU 容量利用不足。AMD Pensando™ Vulcano 800 AI NIC 旨在跨越这一拓扑保持高带宽、低延迟连接,使跨域扩展网络能够在数据中心间有效运行,而无需为集群层次结构的每一层分别配置专用互连。
这两种模式共同构成了现代 AI 集群的完整拓扑。AMD Pensando Vulcano 800 AI NIC 旨在同时满足二者,而解耦式推理和多节点训练已使这一架构要求变得不可或缺。
性能
AI 集群的性能并非仅由峰值吞吐量决定,实际运行条件下的作业完成速度、容错能力和持续利用率同样至关重要。AMD Pensando™ Vulcano 800 AI NIC 旨在全面提升这三项能力。
AMD Pensando Vulcano 800 AI NIC 可将 AI 作业完成速度提升高达 13%。¹ 这一数据反映了减少重传、停顿和恢复周期的潜力,而这些因素会拖慢能力较弱的网络架构中的作业完成速度。在大规模部署中,这一改进会产生复合效应:更快完成作业可直接转化为更高的集群利用率,并在单位运营成本下实现更多训练迭代。
这一性能提升由一系列 RAS(可靠性、可用性和可维护性)能力提供支撑,旨在让集群在真实环境中保持高效运行。多平面架构将流量分布到彼此独立的数据路径上,因此单条链路或单个组件发生故障时,系统性能可平稳降级,而不会中断整个作业。故障隔离、在线诊断和快速恢复机制可缩短平均修复时间,同时无需集群级维护窗口。综合来看,这些能力可提高集群正常运行时间,并维持 AI 基础设施的生产就绪状态。
可编程性
AI 工作负载需求的变化速度已超出传统芯片开发周期所能适应的范围,新的模型架构和通信原语可能在数月内涌现。AMD Pensando™ Vulcano 800 AI NIC 通过完全可编程的软件和硬件堆栈直接应对这一挑战,使协议行为、拥塞控制和传输逻辑能够通过软件进行更新。
这种可编程性建立在对开放标准网络架构的坚定支持之上。AMD 在上一代 AI NIC 中率先将对超以太网联盟(UEC)的支持推向市场,在定义面向 AI 工作负载的下一代以太网方面确立了早期领先地位。AMD Pensando Vulcano 800 AI NIC 延续了这一优势,率先支持 MRC(多路径可靠连接)。MRC 是为 OpenAI 联合开发的一种传输协议,旨在满足大规模训练环境对可靠性和吞吐量的要求。
开放性
基于开放、可互操作标准构建的集群,在市场不断演进时让企业仍能保留供应商选择权并保持灵活性。AMD Pensando™ Vulcano 800 AI NIC 正是针对这种情况而设计的,可在更广泛的以太网生态系统中运行,而无需依赖垂直整合的技术栈来充分发挥能力。
AI 工作负载需求尚未趋于稳定。分布式推理、多节点训练和新架构都在很短的时间内成为标准实践,而且这种演进仍将持续。对基于专有网络架构构建的集群而言,如此快速的变化会造成问题不断累积:每项新需求要么必须落在单一供应商的支持范围内,要么迫使运营方在权宜之计与全面更换网络架构之间作出艰难选择。基于开放标准的基础设施则可完全避免这一限制。由于 AMD Pensando Vulcano 800 AI NIC 原生运行于以太网生态系统中,运营方可以随着需求变化集成新一代硬件并重新配置集群拓扑,而无需受制于供应商的路线图。构建能够跟上工作负载变化步伐的基础设施,其时间窗口正在收窄,而网络架构的选择正处于这一问题的核心。
AI NIC 的具体设计选择进一步提升了这一基础效率。其多平面架构将流量分布到彼此独立的数据路径上,与传统单平面网络架构相比,可使用更少的基础设施实现容错能力和吞吐量。通过减少线缆和收发器的使用,AMD Pensando Vulcano 800 AI NIC 可降低交换成本。
AMD AI NIC™:为 AI 的发展速度而打造
塑造当今 AI 基础设施的各种制约因素,已将网络从辅助考量推向基础性地位。在网络架构层作出的决策,如今会直接影响集群利用率、运营成本,以及随需求变化进行调整的能力。
AMD Pensando™ Vulcano 800 AI NIC 正是基于这一现实打造。无论是横向扩展和跨域扩展连接、作业完成性能、运营韧性、可编程性,还是总体拥有成本,它都旨在满足大规模 AI 部署对网络基础设施提出的特定要求。作为 AMD Helios™ 机架级解决方案的默认网络选择,AMD AI NIC 技术为现代 AI 工作负载所带来的、已无法回避的连接挑战提供了专门构建的解决方案。
附注
- MI400-018:基于 AMD 工程芯片建模和 AMD 综合基准测试模拟,使用 MOE_4p5T hi_sparsity_ea9 基准测试,在采用 FP8 训练数据类型的模拟 LLM 系统上,预测获得解决方案所需时间按天计算的加速幅度。该系统建模采用 8,000 个 AMD Instinct MI455X GPU,并对比每个 GPU 配置 3 个与 2 个 AMD Pensando Vulcano NIC 的情况。结果反映对纯训练计算(仅解码层)的分析。评估配置采用 4,096 的全局批量大小和 4K 序列长度,并使用 SwiGLU 激活函数。结果不包括评估和检查点开销。假设使用具有矩阵乘法与 softmax 重叠的 FlashAttention v3。AllReduce 和 All2All 通信成本均已完整计入,未通过分块计算与通信重叠予以隐藏。在不同重叠程度下评估梯度同步和 FSDP 权重预取,以衡量横向扩展敏感性。结果基于理想且未充分优化的真实环境行为;实际产品上市后的结果可能有所不同。MI400-018
- MI400-018:基于 AMD 工程芯片建模和 AMD 综合基准测试模拟,使用 MOE_4p5T hi_sparsity_ea9 基准测试,在采用 FP8 训练数据类型的模拟 LLM 系统上,预测获得解决方案所需时间按天计算的加速幅度。该系统建模采用 8,000 个 AMD Instinct MI455X GPU,并对比每个 GPU 配置 3 个与 2 个 AMD Pensando Vulcano NIC 的情况。结果反映对纯训练计算(仅解码层)的分析。评估配置采用 4,096 的全局批量大小和 4K 序列长度,并使用 SwiGLU 激活函数。结果不包括评估和检查点开销。假设使用具有矩阵乘法与 softmax 重叠的 FlashAttention v3。AllReduce 和 All2All 通信成本均已完整计入,未通过分块计算与通信重叠予以隐藏。在不同重叠程度下评估梯度同步和 FSDP 权重预取,以衡量横向扩展敏感性。结果基于理想且未充分优化的真实环境行为;实际产品上市后的结果可能有所不同。MI400-018
Related Blogs
附注
- MI400-018: Based on AMD Engineering silicon modeling and AMD synthetic benchmark simulation, using the MOE_4p5T hi_sparsity_ea9 benchmark test to project time- to- solution speed up in days using the FP8 training datatype on a simulated LLM system modeled with 8000 AMD Instinct MI455X GPUs and (3) versus (2) AMD Pensando Vulcano NICs.
Results reflect analysis of pure training compute (decode layers only) Configuration evaluated with a global batch size of 4096 and a sequence length of 4K, using SwiGLU activation. Results exclude evaluation and checkpointing overhead. FlashAttention v3 with matmul–softmax overlap is assumed. AllReduce and All2All communication costs are fully accounted for and not hidden via tiled compute–communication overlap. Gradient synchronization and FSDP weight prefetch are evaluated across varying levels of overlap to assess scale-out sensitivity. Results assume ideal and not fully optimized real-world behavior and may vary when actual product(s) are released in market. MI400-018
- PEN-022: AMD comparison and pricing as of May 18, 2026, for network fabric costs to support 32,000 GPUs. Comparison of a Vulcano-based NIC (VULCANO-CUSTOM-2.4T) deployed as part of a Helios rackscale system with a network based on 1.6T Tomahawk 6 switching with 200G SerDes versus using a competitor 800G NIC with 800G Tomahawk 6 switching with 100G SerDes. Both fabrics were fat-tree topologies built on Tomahawk 5 800G switching platforms, with NIC costs considered comparable. The Vulcano-based design is estimated to deliver up to 33% savings in network switching costs by enabling a more cost-effective architecture with fewer switching platforms, more bandwidth per port on the network, and reduced transceiver cables/optics.
TH6-100G Serdes Fat-Tree (Competition):
Switching (TH6C BCM78914 - 128x800G):
• Leaf Units 1,000
• Spine Units 500
• Total Switches 1,500
• TH6 100G Unit Price $79,587
• Total Switching Cost $60M
Cables/Optics:
• NIC Transceivers (800G-DR8) 64,000 @ $500
• Leaf/Spine Transceivers (800G-DR8) 192,000 @ $500
• MPO Cables 256,000 @ $89
• Total Optics/Cables $64M
Total Fabric Cost TH6-100G (Switches + Cables/Optics): $124M
TH6-200G Serdes Fat-Tree (Vulcano-Custom2.4T / AMD Solution):
Switching (TH6P BCM78910 - 64x1.6T):
• Leaf Units 1,000
• Spine Units 500
• Total Switches 1,500
• TH6 200G Unit Price $66,336
• Total Switching Cost $50M
Cables/Optics:
NIC Transceivers (1.6T-DR8) 32,000 @ $900 (50% fewer vs. competition)
Leaf/Spine Transceivers (1.6T-DR8) 64,000 @ $900
MPO Cables 128,000 @ $89 (50% fewer vs. competition)
Total Optics/Cables $43M
Total Fabric Cost TH6-200G (Switches + Cables/Optics): $93M
Capex Savings (Fabric only):
Savings $: $30.7M
Savings %: 33.1%
Pricing sources: SemiAnalysis Hyperscaler Networking Model data used with permission; full analysis available via SemiAnalysis subscription. Edgecore switch pricing as of May 17, 2026. Results may vary based on system configuration.
- MI400-018: Based on AMD Engineering silicon modeling and AMD synthetic benchmark simulation, using the MOE_4p5T hi_sparsity_ea9 benchmark test to project time- to- solution speed up in days using the FP8 training datatype on a simulated LLM system modeled with 8000 AMD Instinct MI455X GPUs and (3) versus (2) AMD Pensando Vulcano NICs.
Results reflect analysis of pure training compute (decode layers only) Configuration evaluated with a global batch size of 4096 and a sequence length of 4K, using SwiGLU activation. Results exclude evaluation and checkpointing overhead. FlashAttention v3 with matmul–softmax overlap is assumed. AllReduce and All2All communication costs are fully accounted for and not hidden via tiled compute–communication overlap. Gradient synchronization and FSDP weight prefetch are evaluated across varying levels of overlap to assess scale-out sensitivity. Results assume ideal and not fully optimized real-world behavior and may vary when actual product(s) are released in market. MI400-018 - PEN-022: AMD comparison and pricing as of May 18, 2026, for network fabric costs to support 32,000 GPUs. Comparison of a Vulcano-based NIC (VULCANO-CUSTOM-2.4T) deployed as part of a Helios rackscale system with a network based on 1.6T Tomahawk 6 switching with 200G SerDes versus using a competitor 800G NIC with 800G Tomahawk 6 switching with 100G SerDes. Both fabrics were fat-tree topologies built on Tomahawk 5 800G switching platforms, with NIC costs considered comparable. The Vulcano-based design is estimated to deliver up to 33% savings in network switching costs by enabling a more cost-effective architecture with fewer switching platforms, more bandwidth per port on the network, and reduced transceiver cables/optics.
TH6-100G Serdes Fat-Tree (Competition):
Switching (TH6C BCM78914 - 128x800G):
• Leaf Units 1,000
• Spine Units 500
• Total Switches 1,500
• TH6 100G Unit Price $79,587
• Total Switching Cost $60M
Cables/Optics:
• NIC Transceivers (800G-DR8) 64,000 @ $500
• Leaf/Spine Transceivers (800G-DR8) 192,000 @ $500
• MPO Cables 256,000 @ $89
• Total Optics/Cables $64M
Total Fabric Cost TH6-100G (Switches + Cables/Optics): $124M
TH6-200G Serdes Fat-Tree (Vulcano-Custom2.4T / AMD Solution):
Switching (TH6P BCM78910 - 64x1.6T):
• Leaf Units 1,000
• Spine Units 500
• Total Switches 1,500
• TH6 200G Unit Price $66,336
• Total Switching Cost $50M
Cables/Optics:
NIC Transceivers (1.6T-DR8) 32,000 @ $900 (50% fewer vs. competition)
Leaf/Spine Transceivers (1.6T-DR8) 64,000 @ $900
MPO Cables 128,000 @ $89 (50% fewer vs. competition)
Total Optics/Cables $43M
Total Fabric Cost TH6-200G (Switches + Cables/Optics): $93M
Capex Savings (Fabric only):
Savings $: $30.7M
Savings %: 33.1%
Pricing sources: SemiAnalysis Hyperscaler Networking Model data used with permission; full analysis available via SemiAnalysis subscription. Edgecore switch pricing as of May 17, 2026. Results may vary based on system configuration.