How Industry Collaboration Fosters NVIDIA Co-Packaged Optics

TL;DR

• NVIDIA is reshaping data-center connectivity by co-packaging optical and electrical components through deep industry collaboration, forming an ecosystem that spans photonics ICs, electronic ICs, fiber, packaging, connectors, and lasers.
• The Micro Ring Modulator silicon photonics engine enables direct 200Gbps PAM4 modulation per wavelength with a small footprint, unlocking dense, thermally manageable packaging for hyperscale deployments.
• Collaboration with TSMC addresses production-scale challenges, delivering reliable, repeatable performance even at the smallest geometries.
• Quantum-X Photonics and Spectrum-X Ethernet Photonics switches integrate modular optical subassemblies and high-density silicon photonics engines to deliver multi-terabit throughputs and modular laser sources.
• The External Laser Source (ELS) module centralizes light generation, reducing laser count by up to fourfold and enabling field-replaceable maintenance in thermally controlled environments.

Context and background

NVIDIA is advancing data-center connectivity by integrating optical and electrical components and emphasizing collaboration across the industry. The NVIDIA networking platform assembles best-in-class technologies from partners specializing in advanced foundry processes, lasers, and fiber-to-chip solutions. By coordinating each aspect of the solution design, NVIDIA seeks scalable, reliable, and high-performance optical systems that address growing data throughput and energy efficiency needs. The ecosystem encompasses photonics ICs, electronic ICs, fiber, packaging, connectors, and lasers, all designed to support scalable, energy-efficient AI factory interconnects. The efforts center on creating a cohesive, reliable suite of co-packaged optics (CPO) components that can scale with hyperscale data centers. The article highlights how this collaboration underpins the core innovation of CPO—the Micro Ring Modulator silicon photonics engine—which enables high density and power-efficient interconnects. It builds on a cooperative approach with TSMC to tackle manufacturing challenges relevant to micro-scale modulators, aiming for reliable, repeatable performance across production volumes. For context, wafer-level integration of micro-lenses, tight parasitic management, and dense stacking of CMOS and photonics components enable ultra-dense, low-power modules suitable for hyperscale environments. The program also introduces a modular approach to light delivery via external laser sources and modular subassemblies that enable rapid deployment and robust maintenance. This article serves as a focused overview of the innovation, partnerships, and technical foundations for the NVIDIA CPO platform, including photonics ICs, electronic ICs, fiber, packaging, connectors, and lasers. It details how the NVIDIA ecosystem enables scalable and energy-efficient AI factory interconnects, illustrating the breadth of integration—from silicon photonics to system-level packaging—and highlighting the role of industry collaboration in achieving reliable, high-throughput optical interconnects. A key highlight is the Micro Ring Modulator, whose small footprint and thermal stability redefine floorplan constraints and enable multi-row scaling within a single package. From a collaboration standpoint, NVIDIA emphasizes that co-packaged optics are not built in isolation. Deep collaboration with ecosystem partners helps address manufacturing, reliability, and integration challenges, ensuring consistent laser performance and robust packaging at scale. The article positionally frames the NVIDIA ecosystem as a modular, scalable solution that aligns with hyperscale data centers and next-generation AI workloads.

For reference, NVIDIA points readers toward prior CPO technical content and a video deep dive into co-packaged optics switches, as well as NVIDIA’s silicon photonics website for ongoing updates. The exact URLs are provided in the References section.

What’s new

• The Micro Ring Modulator silicon photonics engine is profiled as a breakthrough that combines a small footprint with robust thermal control, enabling extremely dense interconnects without compromising power or thermal budgets. This design enables multi-row scaling of optical interconnect throughput within a single package.
• Production-scale collaboration with TSMC addresses manufacturing problems at small geometries, including precise fabrication control, thermal sensitivity mitigation, and ensuring reliable high-speed modulation across production volumes.
• The Quantum-X Photonics switch introduces an optical subassembly capable of 4.8Tbps transmit and 4.8Tbps receive bandwidth per subassembly, anchored by three COUPE-based optical engines delivering 1.6Tbps transmit and 1.6Tbps receive per engine.
• Each optical engine utilizes eight 200Gbps PAM4 lanes for both transmit and receive, with eight transmit channels, eight receive channels, and two laser input fibers per engine, enabling low-latency, high-throughput communication between the switch fabric and optical interfaces.
• A socket-based subassembly design enables modular connectivity from the electrical interface to the main switch package, plus a hermetically sealed fiber interface to the optical engines for long-term reliability in field deployments.
• The Quantum-X switch ASIC delivers 28.8Tbps full duplex bandwidth when six high-capacity optical subassemblies are integrated on the switch package interposer, showcasing a tightly integrated, thermally efficient, densely packed architecture.
• The Quantum-X Photonics and Spectrum-X Ethernet Photonics platforms are built around an external laser source (ELS) module, a high-power, field-replaceable unit housed in a thermally controlled environment separate from the main chassis, centralizing laser power and reducing total laser counts.
• The ELS module houses eight lasers and can power 32 of the Quantum-X switch’s 576 transmit lanes, enabling a fourfold reduction in laser count and enhancing reliability through centralized light generation. The same ELS concept is applied to Spectrum-X Ethernet Photonics with modular deployment (16 modules for the single-ASIC version and 64 modules for the quad-ASIC version).
• The Spectrum-X Ethernet Photonics package showcases 32 silicon photonics engines within a compact footprint, each engine with 16 transmit and 16 receive lanes (3.2Tbps per engine), and supports 512 lanes of 200Gbps electrical throughput in a small package. A detachable optical connector enables automated, mass-manufacturing workflows and improved assembly yield.
• The architecture is designed to scale toward AI factories, with dense, thermally managed interconnects and robust laser performance across the ecosystem, offering a path to higher bandwidths and lower power per bit.

Why it matters (impact for developers/enterprises)

• Density and energy efficiency: The Micro Ring Modulator’s small footprint and strong thermal control enable multi-row interconnects in a single package, increasing bandwidth per device without escalating power and cooling requirements.
• Reliability and maintenance at scale: Centralized, field-replaceable ELS units reduce the total number of lasers required in the data center, lowering capital and operating costs while improving reliability and serviceability.
• Modular, scalable architecture: Socket-based subassemblies and hermetically sealed fiber interfaces simplify deployment, maintenance, and field upgrades, aligning with hyperscale data-center workflows and automated manufacturing.
• High-throughput AI-ready interconnects: The Quantum-X and Spectrum-X platforms push multi-terabit per second figures per switch, enabling dense, low-latency connectivity for AI inference, training, and other data-center workloads.
• End-to-end collaboration: NVIDIA emphasizes ecosystem partnerships across foundry processes, lasers, and fiber-to-chip solutions, underscoring a holistic approach to co-packaged optics that combines hardware, process control, and system-level packaging for scalable deployment.

Technical details or Implementation

• Micro Ring Modulator: silicon photonics engine engineered for direct 200Gbps PAM4 modulation per wavelength, delivering ultra-low density with high-speed performance and tight thermal control to support hyperscale deployments. This enables dense arrays of high-speed interconnects in compact packages.
• Wafer-level integration: optical modules leverage wafer-level micro-lens integration to streamline fiber alignment, reducing production time and improving scalability.
• COUPE-based optical engines: each optical engine provides 1.6Tbps transmit and 1.6Tbps receive throughput, powered by eight 200Gbps PAM4 lanes for both transmit and receive, with eight transmit, eight receive, and two laser input fibers per engine.
• Quantum-X Photonics subassembly: supports 4.8Tbps transmit and 4.8Tbps receive bandwidth per subassembly; three COUPE-based optical engines per subassembly enable dense co-packaging with high power and thermal efficiency.
• Six subassemblies per switch ASIC: delivering 28.8Tbps full duplex bandwidth when integrated on the switch package interposer, enabling dense electrical and thermal coupling with the optical modules and enabling efficient liquid cooling for both switch core and optical modules.
• Q3450 system: Quantum-X Photonics liquid-cooled switch system delivering 115.2Tbps full-duplex bandwidth over 144 ports with 800Gbps per port, designed for low latency and scalable NVIDIA AI factories.
• Spectrum-X Ethernet Photonics: multi-chip module (MCM) package housing 32 silicon photonics engines; each engine provides 3.2Tbps (16 transmit + 16 receive). The module supports 512 lanes of 200Gbps electrical throughput and uses a detachable optical connector to streamline assembly yield.
• ELS (External Laser Source) module: high-power, field-replaceable laser source housed in a dedicated, thermally controlled environment to stabilize wavelength and aging effects. An ELS module contains eight lasers powering 32 of the Quantum-X switch’s 576 transmit lanes, reducing the total laser count by a factor of four versus legacy designs. The same ELS concept is used for Spectrum-X with 16 modules for single-ASIC and 64 modules for quad-ASIC versions.
• Modularity and field deployment: both Quantum-X and Spectrum-X employ modular ELS units with socket-based subassemblies and hermetically sealed fiber interfaces, enabling rapid deployment, fault isolation, and reliable operation in data-center environments.
• Assembly and reliability: wafer-level integration of micro-lenses, hermetically sealed interfaces, and a detachable optical connector collectively improve production yield, reliability, and maintenance efficiency for hyperscale data centers. | Component | Throughput / Capacity | Notes |---|---|---| | Micro Ring Modulator | 200 Gbps PAM4 per wavelength | Ultra-compact photonics engine with strong thermal control |Quantum-X subassembly | 4.8 Tbps transmit, 4.8 Tbps receive | Core backbone of ultra-high-capacity data paths |COUPE-based optical engines | 1.6 Tbps transmit, 1.6 Tbps receive | Eight 200 Gbps PAM4 lanes per engine; 8 TX, 8 RX, 2 laser inputs |Switch ASIC (per six subassemblies) | 28.8 Tbps full duplex | Integrated on interposer for dense coupling and cooling |Q3450 system | 115.2 Tbps full-duplex | 144 ports at 800 Gbps each |Spectrum-X engines | 3.2 Tbps per engine | 32 silicon photonics engines; 512 electrical lanes |ELS module | Powers 32 lanes (Quantum-X) | Eight lasers; modular, field-replaceable; reduces laser count |

Key takeaways

• Industry collaboration is central to NVIDIA’s CPO strategy, uniting photonics, electronics, packaging, and lasers to deliver scalable optical interconnects.
• The Micro Ring Modulator provides a path to ultra-dense, energy-efficient data-center interconnects through compact packaging and robust thermal management.
• Centralized ELS modules reduce laser counts and improve reliability, while preserving modularity for field maintenance and upgrades.
• The Quantum-X and Spectrum-X platforms demonstrate multi-terabit per second capabilities within densely packed, thermally managed switch ecosystems suitable for AI factories.
• Wafer-level integration, hermetically sealed fiber interfaces, and detachable connectors help bridge production-ready manufacturing with scalable deployment in hyperscale data centers.

References

• https://developer.nvidia.com/blog/how-industry-collaboration-fosters-nvidia-co-packaged-optics/