共封装光学（CPO）书籍：以光进行扩展，迎接下一波互连浪潮

Scale-outandScale-upCPO,CPOTCOandPowerBudgets,DSPTransceiversvsLPOvsNPOvsCPO,TSMC COUPE,MZMvsMRMvsEAM ModulatorDeepDive,CPO FocusedcompaniesandCPO SupplyChain 横向扩展与纵向扩展CPO，CPO的TCO与功耗预算，DSP收发器vsLPOvsNPOvsCPO，TSMC COUPE，MZMvsMRMvsEAM调制器深度解析，聚焦CPO的公司及CPO供应链 DYLAN PATEL、DANIEL NISHBALL、MYRONXIE及另外3⼈ JAN02,2026 2026年1⽉02⽇∙PAID∙付费 Co-Packaged Optics (CPO) has long promised to transform datacenter connectivity, but it hastaken a long time for the technology to come to market, with tangible deployment-readyproducts only arriving in 2025. In the meantime, pluggable transceivers have kept pace withnetworking requirements and remain the default path thanks to their relative cost-effectiveness, familiarity in deployment, and standards-based interoperability. 共封装光学（CPO）长期以来被认为有望彻底改变数据中⼼的连接⽅式，但该技术真正进⼊市场却花费了相当长的时间，直到 2025 年才出现可实际部署的成品。在此期间，可插拔光模块不断跟上⽹络需求的演进，并凭借相对较低的成本、部署上的成熟度以及基于标准的互操作性，仍然是默认的主流选择。 However the heavy networking demands that come with AI workloads mean that this time isdifferent. The AI networking bandwidth roadmap is such that interconnect speed, range,density and reliability requirements, will soon outpace what transceivers can provide. CPOwill provide some benefit and bring more options to scale-out networking, but it will becentral to scale-up networking.CPO will be the main driver of bandwidth increases in scale-up networking for the latter part of this decade and beyond. 然⽽，AI⼯作负载所带来的⾼强度⽹络需求意味着这⼀次情况不同。AI⽹络带宽的发展路线显⽰，互连的速度、距离、密度和可靠性要求很快就会超过现有收发器所能提供的能⼒。CPO 将带来⼀定益处，并为横向扩展⽹络提供更多选择，但它将在纵向扩展⽹络中占据核⼼地位。在本⼗年后半段以及更长时间内，CPO 将成为纵向扩展⽹络带宽提升的主要驱动⼒。 size is limited to one or two racks at most. It is also increasingly difficult to scale bandwidthover copper. In Rubin, Nvidia will deliver another doubling of bandwidth per copper lanethrough bi-directional SerDes, but doubling bandwidth on copper by developing ever-fasterSerDes is a highly challenging vector of scaling that is a slow grind.CPO can deliver the sameor better bandwidth density and can provide additional vectors for scaling bandwidth, allwhile enabling larger scale-up domains. 当今基于铜互连的 scale-up 解决⽅案（如 NVLink）为每颗 GPU 提供了⾼达 7.2 Tbit/s的巨⼤带宽——在 Rubin 这⼀代中即将提升⾄每 GPU 14.4 Tbit/s——但基于铜的链路在距离上最多只能达到两⽶，这意味着 scale-up 域的整体规模最多仅限于⼀到两个机架。此外，在铜介质上继续扩展带宽也正变得越来越困难。在 Rubin 中，Nvidia 将通过双向 SerDes 再次实现每条铜通道带宽翻倍，但依靠开发越来越快的 SerDes 来在铜介质上实现带宽翻倍，是⼀条极具挑战性的扩展路径，推进速度缓慢。CPO 能够提供相同或更⾼的带宽密度，并带来额外的带宽扩展向量，同时还能够⽀持更⼤规模的scale-up 域。 A starting point for understanding the impetus for CPO is to consider the many inefficienciesand trade-offs when using a transceiver for optical communication. Transceivers can be usedto achieve greater link range, but the cage on the front panel of a networking switch orcompute tray that transceivers plug into is typically situated 15-30cm from an XPU or switchASIC. This means that signals mustfirst be transmitted electrically using an LR SerDes overthat 15-30cm distance, with the electrical signal recovered and conditioned by a Digital SignalProcessor (DSP) within the transceiver before being converted into an optical signal. WithCPO, optical engines are instead placed next to XPUs or Switch ASICs, meaning that the DSPcan be eliminated and that lower power SerDes can be used to move data from the XPU tothe Optical Engine. This can reduce energy required to transmit data by more than 50% whencompared to DSP Transceivers - with many aspiring to reduce energy requirements per bit byas much as 80%. 理解 CPO 推动因素的⼀个起点，是审视在使⽤光通信收发器时存在的诸多低效之处和权衡。收发器可以⽤于实现更长的链路距离，但⽹络交换机或计算托盘前⾯板上⽤于插⼊收发器的笼位，通常距离 XPU 或交换 ASIC 有 15–30 厘⽶。这意味着，信号必须先通过 LR SerDes 以电信号形式在这 15–30 厘⽶的距离上传输，并在收发器内部由数字信号处理器（DSP）对电信号进⾏恢复和均衡，然后再转换为光信号。采⽤CPO时，光引擎被直接放置在 XPU 或交换 ASIC 旁边，这意味着可以消除 DSP，并使⽤更低功耗的 SerDes 将数据从 XPU 传输到光引擎。与基于 DSP 的收发器相⽐，这种⽅式可以将数据传输所需的能耗降低 50% 以上，许多⽅案的⽬标甚⾄是将单位⽐特的能耗降低多达 80%。 While scale-out CPO solutions like Nvidia and Broadcom’s are garnering more attention andare being closely looked at by end customers, major Hyperscalers are already starting to planout their scale-up CPO strategy and are committing to suppliers. For instance,Celestial AI isestimating that they could generate a $1B revenue run rate by the end of calendar year 2028-we believe this will primarily be driven by a CPO scale-up solution shipping with Amazon’sTrainium 4. 尽管像英伟达（Nvidia）和博通（Broadcom）这样的横向扩展（scale-out）CPO 解决⽅案正获得更多关注，并被终端客户密切评估，但主要的超⼤规模云服务商已经开始规划其纵向扩展（scale-up）CPO 战略，并向供应商作出承诺。例如，Celestial AI 估计，到 2028⽇历年末，其可实现 10 亿美元的收⼊年化规模——我们认为，这主要将由随亚马逊 Trainium 4⼀同出货的 CPO 纵向扩展解决⽅案所驱动。 CPO focused companies are now well beyond papers, pilot projects and demonstrations andand are making key product decisions such as optical port architecture to solve for highvolume manufacturing. CPO for scale-up is nownot a matter of if and why, but when andhow– how to bring these systems to volume production, and when key component supplychain companies like laser manufacturers can ramp up sufficient production. 专注于 CPO 的公司如今早已超越论⽂、试点项⽬和演⽰阶段，正在为⾼产量制造做出诸如光端⼜架构等关键产品决策。⽤于规模化的 CPO 已不再是是否或为何的问题，⽽是何时以及如何的问题——如何将这些系统推向量产，以及诸如激光器制造商等关键元器件供应链企