報告_OCP2026_Micas_LPO到NPO到CPO路線_20261015

PDF 原檔：報告_OCP2026_Micas_LPO到NPO到CPO路線_20261015_original.pdf

講者公開簡報（OCP 官方 Recorded Content 頁提供下載）。文字以 PyMuPDF 逐頁抽取，### p<n> 為頁碼；圖表內容見原檔。

圖片清單（已驗證 2026-10-19）

原始內容

p1

From LPO to NPO to CPO The open optical roadmap for AI infrastructure A staged architecture view for pluggable optics, linear pluggable optics, near- package optics, co-packaged optics, and co-packaged copper. C O R E T H ES I S Optical I/O will move closer to silicon. The real question is timing, service model, and production readiness. Tim Zhou, VP, Micas Networks Monday, October 12 · 12:30 PM

p2

A G E N D A Where Micas is going today 01 Thesis p. 03 Direction is clear; timing is the open question. 02 Architecture options p. 04–05 Pluggable, LPO, NPO and CPO, and the problem each one solves. 03 NPO up close p. 06–07 Three NPO paths and the laser choice that shapes service. 04 Standards & outlook p. 08–10 Three new MSAs and what 2026–2029 looks like. 05 Execution risks p. 11 The bottlenecks that set the real ramp. 06 Micas' role & call to action p. 12–13 What we bring, and four asks for the OCP community. OCP executive session · Open optical roadmap 02

p3

C O R E T H E S I S Direction is clear. Timing is not. 01 Direction AI cluster growth forces optical I/O closer to switch ASICs and accelerator packages. 02 Timing The ramp depends on manufacturing maturity, qualification cycles, thermal design, and service risk. 03 Open ecosystem Hyperscalers want to turn proprietary systems into reusable interfaces, test methods, and operational norms. 2026 Early adoption of CPO and supply-chain integration year. 2027 Highest uncertainty: readiness varies by use case. 2028 – 2030 XPO/NPO/CPO co-existence? OCP executive session · Open optical roadmap 03

p4

A R C H I T E C T U R E R O L E S Pluggables remain necessary, but not sufficient closer to the ASIC → Pluggable 800G / 1.6T LPO NPO CPO PRIMARY VALUE Operational maturity and broad ecosystem Lower module power and latency by reducing DSP dependency Shorter electrical reach with better serviceability than CPO Best density and power efficiency close to the ASIC MAIN TRADE-OFF Higher electrical loss and power at high lane rates Tighter host-channel and link- margin requirements More integration complexity than pluggable optics System yield, thermal, repair, and field-service complexity BEST-FIT WINDOW NOW & ONGOING Immediate and continuing volume NEAR-TERM AI fabrics with validated SI margin 2026 – 2028 Platform transition POST-QUALIFICATION High-density systems The practical roadmap is not a single replacement cycle. It is a portfolio decision by distance, density, power, latency, cost, and service model. OCP executive session · Open optical roadmap 04

p5

C O M P L E M E N T A R Y S T A C K CPO, LPO and NPO solve different problems CO -PA CK AG ED OP TI CS CPO Maximum integration for power and density. Shortens the ASIC-to-optics electrical channel, but changes the failure boundary and requires strong yield, rework, thermal, and service strategies. LI NE AR PL UG GA BLE O PT ICS LPO Practical near-term power reduction. Reduces module-side electronics and latency, but requires careful host-channel characterization, port selection, and platform-specific validation. NE AR -PA CK AG E O PT IC S NPO Serviceable optical proximity before full CPO. Shortens the electrical reach while preserving a clearer optical-engine replacement model. The key shift for AI infrastructure is not bandwidth alone. It is bandwidth… bandwidth per watt bandwidth per rack unit bandwidth per dollar bandwidth per operational risk OCP executive session · Open optical roadmap 05

p6

S E R V I C E A B L E O P T I C A L P R O X I M I T Y NPO deep dive NPO moves the optical engine close enough to reduce 200G/lane electrical reach while keeping a practical manufacturing and repair boundary. NPO PATH HOW IT CONNECTS TECHNICAL VALUE ENGINEERING RISK Open CPX-based NPO ASIC → CPX socket → serviceable optical engine Open optical-engine boundary with stronger vendor flexibility Socket, power, cooling, telemetry, firmware, and OE qualification OBO / OCI path Traditional BGA ASIC with very short PCB path to modular OBO optical engine Faster platform introduction with independent optical subsystem qualification Connector loss, external laser distribution, CMIS management, and production test NPO reduces CPO adoption risk only when the system vendor owns the full architecture: SI, thermals, mechanics, software, telemetry, test, and field service. OCP executive session · Open optical roadmap 06

p7

N P O D E S I G N C H O I C E Laser architecture shapes the NPO service model External laser source + ADVANTAGES Keeps laser heat and aging outside the OE; enables laser redundancy and replacement. − DISADVANTAGES Adds light-source modules, fibers, connectors, routing, cleanliness burden, and management objects. BEST FIT High-density CPO or NPO where serviceable laser replacement and redundancy matter most. Internal laser in NPO module + ADVANTAGES Simplifies optical plumbing and creates a self-contained field- replaceable assembly. − DISADVANTAGES Places laser heat, aging, monitoring, and failure impact inside the OE module. BEST FIT Serviceable NPO where customers want built-in-laser modules and less external laser infrastructure. H Y P E R S C AL E R R E Q U I R E M E NT S I G N A L Built-in laser NPO can simplify service procedure and reduce dependency on external light-source plumbing. M I C A S E VA L U A T I O N C O N T EX T Track various NPO system approaches, including built-in-laser module concepts and optical chipset ecosystem options. OCP executive session · Open optical roadmap 07

p8

I N D U S T R Y S T A N D A R D S 3 new MSAs 08

p9

T H R E E - Y E A R V I E W · 1 O F 2 2026 – 2029 outlook: pluggable paths TECHNOLOGY SCALE-OUT ROLE SCALE-UP ROLE 2026 – 2029 OUTLOOK KEY CHALLENGES Pluggable Mainstream leaf-spine and backend Ethernet fabric Very limited role where reach and serviceability matter more than density Continues through 2026–2029 as the volume baseline Power, front-panel density, long PCB channel at 200G/lane LPO Near-term lower-power option for AI Ethernet scale-out Possible for short controlled links, but margin-sensitive 2026 validation and selective deployment; broader in 2027 only where SI margin is proven Host channel loss, connector quality, SerDes tuning, thermal drift, vendor interoperability XPO Higher-density pluggable path between LPO and true NPO Useful where serviceable dense optics are preferred over package-adjacent integration 2027 engineering exploration; may coexist with NPO for operators needing a pluggable service model Still has more electrical reach than NPO/CPO; connector and module ecosystem maturity OCP executive session · Open optical roadmap 09

p10

T H R E E - Y E A R V I E W · 2 O F 2 2026 – 2029 outlook: near- & co-packaged TECHNOLOGY SCALE-OUT ROLE SCALE-UP ROLE 2026 – 2029 OUTLOOK KEY CHALLENGES OCI MSA / OBO optical engine Strong fit for scale-out switch ASICs using near- board optical engines Possible optical scale-up building block where OCI- style links are adopted 2026–2028 early systems around TH6- class platforms; expands as OCI ecosystem matures External laser distribution, OE qualification, fiber routing, CMIS/software model Open CPX NPO · socket Serviceable socketed OE near switch ASIC or host device Strong candidate for modular optical scale-up if xPUs adopt CPX sockets 2027 prototype and ecosystem buildout; 2028 early deployment if interoperability improves Socket standard execution, power delivery, cold plate, CMIS behavior, interchange rules CPO Best long-term option for highest radix and lowest power per bit Later-stage option for XPU or accelerator package integration 2026 validation and selective deployment. Switch CPO grows first in 2027–2029; XPU-CPO follows later System yield, repair boundary, external laser reliability, fiber attach, thermal, field service No single interconnect will replace the others. The market will segment by reach, density, power, latency, cost, maintainability, software readiness, and manufacturing maturity. OCP executive session · Open optical roadmap 10

p11

E X E C U T I O N R I S K S Common bottlenecks determine the real ramp 01 Electrical channel 200G/lane and future 400G/lane margins depend on connector transitions, return loss, crosstalk, skew, package escape, and low-loss materials. 02 Thermal/Power design CPO and NPO place temperature-sensitive optics near 1–2 kW class ASIC packages, requiring reliable cold plates, TIMs, leak detection, and coupled control. 03 Yield & serviceability OE, fiber coupling, package interface, and laser failures must not turn every issue into expensive ASIC or full switch replacement. 04 Fiber & connectors High fiber count drives bend radius, routing, contamination control, automated assembly, density, and field repair complexity. 05 Software & operations SONiC and platform management must handle OE bank mapping, CMIS, shared lasers, coordinated state changes, firmware lifecycle, and fault isolation. The ramp will be won through production engineering, not headline architecture labels. OCP executive session · Open optical roadmap 11

p12

S Y S T E M S U P P L I E R C O N T R I B U T I O N Micas' role in an open optical ecosystem FOCUS AREA TECHNICAL CAPABILITY CUSTOMER VALUE CPO platforms CPO system integration, ELS control, fiber routing, thermal design, diagnostics, and SONiC integration Lower power per bit and higher density once qualification and service models are ready LPO enablement Host-channel characterization, port selection, SI/PI validation, module interoperability, and thermal margin testing Near-term power reduction without waiting for full CPO deployment NPO readiness Near-package architecture planning, short-reach electrical design, OE placement, maintainability analysis, and software abstraction Lower-risk transition between pluggable optics and fully integrated CPO SONiC & platform software Device drivers, CMIS management, DOM telemetry, firmware workflows, port mapping, diagnostics, and automation Consistent operating model across pluggable, LPO, NPO, and CPO Open collaboration matters because customers need interchangeable learning across interfaces, optics, cooling, management, and manufacturing test. OCP executive session · Open optical roadmap 12

p13

O P E N E X E C U T I O N A G E N D A OCP call to action 1 Open interface definition Align CPX, OCI/OBO, ICA-based NPO, CPO, and CPC around common electrical, thermal, power, and service boundaries. 2 Common validation methods Standardize SI, BER, thermal, laser, firmware, and field-service test practices. 3 Open software control Expose optical engine state, laser health, alarms, firmware, and recovery workflows through SONiC and platform management. 4 Interoperable supply chain Enable multi-vendor optical engines, sockets, lasers, connectors, cooling, and manufacturing test ecosystems. CPO adoption may be staged, but the migration of optics closer to silicon is inevitable. The open ecosystem should now focus on production readiness at scale. We opened with one thesis: optics move closer to silicon. The question was never if, only timing and readiness. OCP executive session · Open optical roadmap 13

p14

Thank you. Tim Zhou VP, Micas Networks JO IN U S W ED NE SDA Y · RSV P Open AI Factory After Hours Wed, Oct 14 · 6:00 – 9:30 PM 367 S 1st St, San Jose Food, drinks and the newest CPO from Micas and Broadcom. Scan to RSVP Booth E5 ST OP B Y T HE B OOT H Enter the Deebo raffle Win a framed, signed Deebo Samuel photo. Visit Micas at booth E5 for a live CPO demo and enter to win. Scan to enter OCP executive session · Open optical roadmap 14