MS Input – One Page Summary
(For Microsoft Review / Please Forward to Microsoft)
IB Rack-Level Optical Shuffle System (MPO8, W80S0)
Purpose
This note is intended to allow MS to assess whether the system architecture and delivery model described below are suitable for InfiniBand rack deployments, and not to validate individual product certifications.
System Overview
We propose a rack-level optical shuffle system for InfiniBand aggregation racks in hyperscale data centers.
The system is designed to standardize high-fiber-count optical aggregation and re-mapping within a single rack,
and to serve as a repeatable rack baseline once qualified.
This is not a one-off component selection.
What This Is
- Rack-level OS2 single-mode MPO8 optical shuffle system
- Delivered as a complete kit, not loose cables
- Fixed polarity architecture (Type A only)
- High fiber-count trunks (96F / 144F / 288F)
- Intended for IB aggregation racks
The kit includes:
- Pre-terminated MPO8 trunks
- Adapter plates
- Cable management components
- 100% IL / RL optical testing
- Cable-level labeling and full traceability
Why This Matters
In IB racks, optical shuffle is a high-risk failure point:
- Polarity errors impact the entire rack
- Inconsistent insertion loss affects link stability
- Poor labeling complicates field troubleshooting
The objective is to control these risks at the rack-system level, rather than managing individual cables ad hoc.
Key Technical Constraints
- OS2 single-mode fiber
- MPO8 connectors
- Type A polarity only
- 100% optical testing (no sampling)
- Serial-number–linked test data and traceability
- Rack-level kit delivery for consistent installation
NVIDIA Reference Design Alignment
This system aligns with NVIDIA InfiniBand Reference Design at the system architecture level.
Passive MPO cabling products are not NVIDIA-certified items.
The intent is to ensure that the optical architecture, polarity model, testing depth, and delivery discipline
are consistent with NVIDIA-based IB deployments commonly used in hyperscale data centers.
Question for MS
Based on your experience with IB racks:
- Is this rack-level MPO8 shuffle architecture a familiar and acceptable approach in MS IB deployments?
- Are there any red flags related to polarity model, testing depth, or rack-level delivery?
End of Summary
MS 输入版 · 一页说明
(供 Microsoft 审阅|请转发至 Microsoft)
IB 机柜级光纤互连与汇接系统(MPO8,W80S0)
目的
本说明用于协助 MS 判断下述 系统架构与交付模型 是否适用于 InfiniBand 机柜部署,
并非用于验证单个被动产品的认证情况。
系统概述
我们提出的是一套面向 Hyperscaler 数据中心 InfiniBand 聚合机柜的
机柜级光纤互连与汇接系统。
该系统用于在单一机柜内部实现高纤芯数光纤的标准化聚合与重映射,
并在完成验证后,作为 可重复使用的机柜级基线配置进行部署。
这并不是一次性的零部件选型。
系统定义
- 机柜级 OS2 单模 MPO8 光纤互连与汇接系统
- 以完整 Kit 形式交付,而非零散线缆
- 固定极性架构(仅 Type A)
- 高纤芯数 Trunk(96F / 144F / 288F)
- 面向 InfiniBand 聚合机柜 应用场景
该 Kit 包含:
- 预端接 MPO8 Trunk 光纤
- MPO 适配板
- 线缆管理组件
- 100% 插损 / 回损光学测试
- 线缆级标签与完整可追溯体系
为什么重要
在 InfiniBand 机柜中,光纤互连与汇接是一个 高风险失效点:
- 极性错误可能影响整柜链路
- 插损一致性不足会影响链路稳定性
- 标识不清会显著增加现场排障难度
该系统的目标,是在机柜系统层面对上述风险进行控制,
而不是以零散线缆方式被动管理。
关键技术约束
- OS2 单模光纤
- MPO8 连接器
- 仅 Type A 极性
- 100% 光学测试(不允许抽检)
- 测试数据与序列号绑定,实现可追溯
- 以机柜级 Kit 方式交付,保证安装一致性
NVIDIA 参考设计对齐说明
该系统在系统架构层面对齐 NVIDIA InfiniBand Reference Design。
被动 MPO 光纤产品本身并不存在 NVIDIA 官方的单独产品认证。
本方案的核心在于确保:
- 光纤架构设计
- 极性模型
- 测试深度
- 交付与管控方式
与 Hyperscaler 数据中心中常见的 NVIDIA IB 部署实践保持一致。
给 MS 的问题
基于您在 InfiniBand 机柜部署方面的经验:
- 这种 机柜级 MPO8 光纤互连与汇接架构 是否为 MS 当前部署中熟悉且可接受的方式?
- 在极性模型、测试深度或机柜级交付方式上,是否存在明显风险点或需要调整之处?
—— 完 ——
Concept Clarification and MS Assessment Focus
NVIDIA Certification Context
NVIDIA does not certify passive MPO cabling products at the individual item level.
In practice, data center projects follow NVIDIA InfiniBand Reference Design at the system architecture level, together with each hyperscaler’s own qualification and approval processes.
What Is Being Proposed
What we are proposing is a rack-level MPO8 optical shuffle system, aligned with NVIDIA InfiniBand reference architecture in terms of:
- Polarity model
- Testing depth
- Delivery and installation discipline
The proposed system is supplied by factories that already participate in NVIDIA-based deployments through Tier-1 or Prime partners.
What We Ask MS to Assess
The focus is not whether a passive cable is “NVIDIA certified”.
Instead, we seek MS guidance on whether:
- This rack-level optical shuffle architecture, and
- This rack-level delivery and supply approach
are consistent with, acceptable to, and low-risk for MS InfiniBand rack practices.
IB Rack Shuffle Kit (W80S0)
Candidate Factory Paths – MS-Friendly Reference
Context Note
The vendors referenced below are not claimed to be NVIDIA-certified passive fiber products.
They represent factory categories and representative suppliers whose MPO-based optical interconnect systems have been practically deployed in NVIDIA-based InfiniBand or hyperscale data center projects, either directly or via Tier-1 / Prime integrators.
The purpose is to allow MS to determine whether this system architecture and these supply paths are familiar and acceptable within MS IB rack deployments.
Path 1 – Highest Priority
Tier-1 Optical Interconnect Vendors / Deep NVIDIA Ecosystem Participants
Characteristics
- Long-term participation in NVIDIA IB / AI data center deployments
- May not market a product explicitly named “Shuffle Kit”
- Proven use of MPO8 trunks and high-fiber-count pre-terminated assemblies inside IB racks
Representative Examples
- US Conec / Senko ecosystem manufacturers (connector + cable assembly)
- Coherent (formerly II-VI) optical interconnect manufacturing groups
- Molex (Optical Interconnect Division)
Path 2 – Most Commonly Accepted in Practice
Tier-2 Optical System Suppliers to Hyperscaler / Tier-1 ODMs
Characteristics
- Not branded as NVIDIA suppliers
- Long-term suppliers to NVIDIA partner ODMs and system integrators
- Capable of 144F / 288F MPO8 trunks, 100% testing, and rack-level kit delivery
Typical Supplier Profiles
- Optical system factories supplying Quanta / Foxconn / Wistron / Inventec
- Prior delivery experience in AI clusters or IB aggregation racks
Path 3 – Supplemental Path (Scale and Cost Balance)
Mature Hyperscale Data Center Cabling System Vendors (Non-NVIDIA-Specific)
Characteristics
- Strong footprint in Microsoft / Meta / Google data center ecosystems
- Mature process control and delivery discipline
- Deep experience with polarity management, testing, and traceability
Representative Examples
- CommScope (SYSTIMAX / NETCONNECT)
- Corning (pre-terminated fiber systems)
- Belden / Panduit (data center optical infrastructure)
Unified Mandatory Requirements (Vendor-Independent)
All candidate factories, regardless of supply path, must meet all of the following:
- OS2 single-mode MPO8
- Type A polarity only (zero tolerance)
- Proven mass production of ≥144F / ≥288F trunks
- 100% IL / RL testing (no sampling)
- Serial-number-level traceability per cable
- Rack-level kit delivery (not loose components)
Qualitative Question for MS
Based on MS experience with InfiniBand racks:
- Does this rack-level MPO8 shuffle system approach, using factories from the above supply paths, align with MS IB deployment practices?
- Are there any concerns related to architecture, polarity model, testing depth, or delivery model?
End of Document
概念澄清与 MS 判断重点
NVIDIA 认证背景说明
NVIDIA 本身并不对 MPO 光纤线束或 Shuffle Kit 等被动布线产品出具单独的产品认证。
在实际数据中心项目中,遵循的是 NVIDIA InfiniBand Reference Design(系统级),并结合各家 Hyperscaler 自身的 qualification 与审批体系。
当前方案的定位
我们提出的是一套 机柜级 MPO8 光纤互连与汇接系统,在以下方面与 NVIDIA IB 参考架构对齐:
- 极性模型
- 测试深度
- 交付与安装方式
该系统采用的工厂路径,均来自 已通过 Tier-1 / Prime 伙伴参与 NVIDIA 体系项目的供应链。
希望 MS 帮助判断的核心问题
我们关注的重点 不是“一根被动光纤是否有 NVIDIA 认证”。
而是希望 MS 判断:
- 这种 机柜级光纤互连与汇接系统架构,以及
- 这种 机柜级交付与供应路径
是否符合 MS 当前的 InfiniBand 机柜实践,是否可接受、可落地、且不存在明显风险。
IB 机柜光纤互连与汇接系统(W80S0)
候选工厂路径清单(MS 参考版)
说明
以下并非宣称“被 NVIDIA 认证的被动光纤产品”,
而是指在 NVIDIA-based InfiniBand / Hyperscaler 数据中心项目中,已经被实际采用或通过 Tier-1 / Prime 路径供货的工厂类型与代表性厂商。
目的在于让 MS 判断:
该系统架构与这些供应路径,是否在 MS 的 IB 机柜体系中是熟悉且可接受的。
路径一(优先级最高)
Tier-1 光互连厂商 / NVIDIA 生态深度参与者
特点
- 长期参与 NVIDIA IB / AI 数据中心项目
- 不一定以 “Shuffle Kit” 名义销售产品
- 其 MPO8 Trunk 与高纤芯预端接组件 已在 IB 机柜中实际使用
代表性厂商
- US Conec / Senko 体系工厂
- Coherent(原 II-VI)光互连制造体系
- Molex(光互连事业部)
路径二(现实可行、最常被接受)
Hyperscaler / Tier-1 ODM 的二级光纤系统供应商
特点
- 不直接挂 NVIDIA 品牌
- 长期为 NVIDIA 合作 ODM / 集成商供货
- 具备 144F / 288F MPO8 Trunk、100% 测试与 Kit 交付能力
典型类型
- 为 Quanta / Foxconn / Wistron / Inventec 提供高密度 MPO 预端接系统
- 具备 AI Cluster / IB 聚合机柜 项目经验
路径三(补充路径,用于规模与成本平衡)
成熟 Hyperscaler 数据中心布线系统厂商(非 NVIDIA 专属)
特点
- 更偏向 Microsoft / Meta / Google 数据中心生态
- 工艺成熟、流程纪律性强
- 对 极性管理、测试、追溯有深度理解
代表性厂商
- CommScope(SYSTIMAX / NETCONNECT)
- Corning(预端接光纤系统)
- Belden / Panduit(数据中心光纤系统)
候选工厂统一硬性要求(与厂商无关)
无论来自哪条路径,必须同时满足:
- OS2 单模 MPO8
- Type A 极性,零容错
- ≥144F / ≥288F 高纤芯 Trunk 量产经验
- 100% IL / RL 测试(非抽检)
- 单根线缆 序列号级可追溯
- 机柜级 Kit 交付(非散件)
给 MS 的定性问题
基于 MS 在 InfiniBand 机柜部署方面的经验:
- 采用上述供应路径的 机柜级 MPO8 光纤互连与汇接系统方案,是否符合 MS 的 IB 机柜实践?
- 在架构、极性模型、测试深度或交付方式上,是否存在明显风险或需要调整之处?