NVIDIA LinkX Cables and Transceivers for High-Speed AI Factory Networking

InfiniBand and Ethernet interconnects for 100G, 200G, 400G, and 800G infrastructures.

Overview Next-Generation Interconnects for NVIDIA AI Factory Networking

Interconnects have become the critical backbone for scaling artificial intelligence clusters. As AI models grow exponentially, AI factories face physical limitations in bandwidth, distance, reliability, and power efficiency.

NVIDIA LinkX™ optical transceivers, cables, and co-packaged optics (CPO) are an integral part of the NVIDIA full stack AI solution, enabling the performance, reliability, and scale required of AI factories across a range of applications, reaches, and speeds. 

LinkX products are supported end-to-end for both NVIDIA Quantum InfiniBand and NVIDIA Spectrum™-X Ethernet architectures for switch-to-switch, server-to-switch, and storage systems. 

All LinkX products are fully designed, qualified, tested, and supported in NVIDIA systems to minimize time to first token and to maximize cluster performance and uptime.

Co-Packaged Silicon Photonic Networking Switches

NVIDIA CPO provides data throughput, low latency, and drastically improved reliability and power efficiency to scale to millions of GPUs in AI factories.

Spectrum-X Ethernet Photonics: The World’s First Scalable Network for AI Factories

NVIDIA Spectrum-X Ethernet Photonics directly impacts application-level AI performance, enhancing the productivity and output of gigascale AI factories.

Benefits

What are the Benefits of Using NVIDIA Optics and Cables?

AI Factory Scale

LinkX enables AI factory-scale networking with optics delivery, capacity, and supply chain resilience designed to support millions of GPUs.

Lower Power

NVIDIA pluggable and co-packaged optics reduce network power, with CPO delivering 5x better power efficiency than off-the-shelf transceivers.

Lower Latency

LinkX optics reduce or eliminate digital signal processing (DSP) retiming, delivering low, consistent latency that improves AI network responsiveness.

Resiliency

NVIDIA optics deliver higher resiliency and fewer failure points, leading to more cluster uptime.

End-to-End Support

LinkX is qualified and tested inside NVIDIA networking systems, giving customers end-to-end performance and support for AI infrastructure.

Faster Deployment

LinkX simplifies installation and serviceability, supports interoperability across systems, and is positioned to deliver faster time to first token.

How do NVIDIA CPO Switches With Silicon Photonics Scale to Millions of GPUs and Multi-Site AI Factories?

Products Optical Transceivers and Cables for NVIDIA Networking

NVIDIA LinkX provides comprehensive, end-to-end interconnect products designed to maximize the performance of NVIDIA Quantum InfiniBand and Spectrum-X Ethernet networking.  

Optical Transceivers

NVIDIA LinkX optical transceivers deliver high-density, low-power, and low-latency fiber interconnects, fully codesigned and validated with NVIDIA Quantum InfiniBand and Spectrum-X Ethernet platforms to ensure seamless interoperability and maximum cluster performance.

Co-Packaged Optics

NVIDIA LinkX co-packaged optics integrate optical engines directly alongside the switch ASIC, eliminating electrical SerDes loss to deliver significantly lower power consumption, reduced latency, and higher bandwidth density than traditional optics. CPO is purpose-built to scale AI supercluster interconnects to 1.6 terabits per second (Tb/s) and beyond.

Linear Active Copper Cables (LACC)

NVIDIA LinkX linear active copper cables integrate active equalization in signal conditioning to extend usable reach over a cost-effective, low-power copper cable. LACCs are optimized for rack-to-rack AI fabric deployments where deterministic, low-latency data movement is critical. 

Passive Direct-Attach Copper Cables (DAC)

NVIDIA LinkX passive direct-attach copper cables provide a direct twinaxial copper connection between ports with no active components. This delivers ultra-low power and latency in fixed and breakout configurations for cost-effective, short-reach, top-of-rack, switch-to-server, and switch-to-switch deployments.

Specifications NVIDIA Optical Transceiver and Cable Specifications

Feature CPO 1.6T/800G OSFP 800G/400G OSFP 400G QSFP-DD
Modulation 200G PAM4 200G PAM4 100G PAM4 50G PAM4
Protocol InfiniBand and Ethernet InfiniBand and Ethernet InfiniBand and Ethernet Ethernet only
Max DAC Reach n/a n/a 3 m 2.5 m
Max LACC Reach n/a 2.5 m 5 m 5 m
Max Optic Reach 500 m (DR)/2 km (FR) 500 m (DR)/2 km (FR) 500 m (DR)/2 km (FR) 500 m (DR)/2 km (FR)
Switch Connector MPO12/MMC16 OSFP OSFP QSFP-DD
Network Interface Card (NIC) Connector OSFP/QSFP112 OSFP/QSFP112 OSFP/QSFP112 QSFP56/QSFP28
Compatible Switches Q3450-LD, SN6800, SN6810 Q3400/3200, SN6600 QM9700, SN5610 SN5400, SN4700
Compatible NICs NVIDIA ConnectX™-9 ConnectX-7/8 ConnectX-7 ConnectX-6/7 

Resources Explore Cables and Transceivers Resources

Accelerate AI with NVIDIA Silicon Photonics-Based Network Switching

Explore why co-packaged silicon photonics can accelerate AI with benefits like lower power consumption, reduced latency, and network resiliency. 

Maximize Better Power Efficiency for AI Factories

Learn how NVIDIA Spectrum-X Ethernet Photonics directly impacts application-level AI performance, enhancing the productivity and output of gigascale AI factories.

Dive Into NVIDIA Co-Packaged Optics-Based Switches

Take a look inside the NVIDIA silicon photonics-based networking switches.

Delve Into Technical Documentation for LinkX

Read more about the industry’s most complete line of 10, 25, 40, 50, 100, 200, 400, and 800GbE in Ethernet products and EDR, HDR, NDR, and XDR in InfiniBand products.

Frequently Asked Questions About NVIDIA LinkX Cables, Transceivers

NVIDIA LinkX is a portfolio of passive and active copper cable, optical transceivers, passive jumper cables, and co-packaged optics (CPO) designed to power high-performance AI factory networking. As an integral part of NVIDIA's full-stack AI solution, LinkX products deliver the bandwidth, reliability, and power efficiency needed to scale modern AI clusters.

LinkX products are supported end-to-end across both NVIDIA Quantum InfiniBand and Spectrum-X Ethernet architectures, making them compatible with a wide range of AI factory network configurations.

LinkX supports switch-to-switch, server-to-switch, and storage system connections, providing comprehensive connectivity across the full AI factory infrastructure.

As AI models grow in scale and complexity, interconnects become the critical backbone of AI clusters. Limitations in bandwidth, distance, reliability, and power efficiency can all bottleneck AI performance. LinkX is purpose-built to address these challenges and keep AI factories running at full capacity at all times.

LinkX products cover a range of applications, reaches, and speeds up to 1.6 terabits per second (Tb/s). For XDR 2x800G and 1.6T Ethernet deployments, copper options (LACC/AEC) reach up to 2.5–3 meters, while single-mode optical transceivers reach up to 500 meters (DR4) and up to 2 kilometers (FR4).

Co-packaged optics (CPO) integrates optical components directly with networking silicon to reduce power consumption and improve signal integrity at high speeds. LinkX includes CPO solutions where silicon photonics modulators are mounted directly on the switch ICs rather than in separate pluggable transceivers.

XDR (eXtreme Data Rate) is NVIDIA’s 800G InfiniBand generation. It’s based on 200G PAM4 (pulse amplitude modulation with four levels) signaling, which encodes 2 data bits per clock pulse at approximately 100 GHz clock rate, doubling the per-lane throughput compared to NDR’s 100G PAM4.

DR4 (data center reach) supports up to 500 meters using four parallel channels over eight fibers with MPO-12/APC connectors. FR4 (far reach) supports up to two kilometers by multiplexing four channels of different laser wavelengths into a single fiber pair using Lucent connectors (LCs). The two types are not interchangeable. DR4 is parallel and FR4 is multiplexed, so they can’t be directly linked to each other.

IHS (integrated heat sink) transceivers have cooling fins built into the plug and are used exclusively in NVIDIA Quantum switches (Q3400, Q3200, and Spectrum SN56x0 Ethernet switches). RHS (riding heat sink) transceivers have a flat-top design and are used in ConnectX-8 adapters, DGX B300, and NVL GB300 compute systems. The two types are not interchangeable due to their different cooling requirements.

Both are active copper cable technologies, but they serve different roles. AECs use digital signal processors (DSPs) at each end to restore and retime signals, enabling reach up to three meters, primarily for switch-to-switch connections, consuming ~27 watts per end. LACCs use lower-power linear amplifier ICs (~4 watts) to extend signal reach up to 2.5 meters and are used for switch-to-adapter (e.g., ConnectX-8) connections.

At XDR’s 200G PAM4 signaling rates, passive DAC cables have a maximum usable length of under 0.7 meters, which is too short to be practical in most deployments. For this reason, all NVIDIA XDR copper cables are active (either LACC or AEC), which extend usable reach to 2.5–3 meters.

The 200G PAM4 XDR signal operates at a ~100 GHz clock rate, compared to ~50 GHz for NDR’s 100G PAM4. The faster signal transitions cause wires to act more like antennas, radiating signal energy off the cable. This signal degradation limits usable copper length, which is why XDR LACCs are shorter than NDR ACCs.

XDR and CPO deployments use the same MPO-12/APC single-mode crossover fiber cables as NDR/400G, meaning customers can reuse existing NDR fiber infrastructure. NVIDIA offers straight and 1:2 splitter MPO-12/APC cables in lengths from three to 150 meters. For reaches beyond 150 meters or LC-LC fiber cables, third-party suppliers such as Corning, Panduit, or CommScope should be used.

XDR DR4-based transceivers use MPO-12/APC (multiple push-on 12-fiber, angled polished connector) connectors. The angled polish reduces back-reflections that can distort data signals. Only eight of the 12 fiber positions are used: four transmit and four receive. FR4-based transceivers (2 km reach) use LC duplex connectors instead.

Yes, with some nuance. The Q3200 switch supports NDR 400G cables and transceivers for backwards compatibility with 100G PAM4 and 50G PAM4 HDR. The Q3400 switch supports only select NDR cables and transceivers. XDR transceivers cannot downshift to 100G PAM4 on their own. Backwards compatibility is handled at the switch or host channel adapter (HCA) level, not in the cables or transceivers.

No. Copper cables (DAC, LACC, and AEC) can’t be used with the Q3450 CPO switch. The CPO switch uses MMC-12/APC fiber connections only.

Next Steps

Networking Configuration Tools for InfiniBand and Ethernet Clusters

NVIDIA provides a series of networking tools to properly configure a cluster based on your choice of interconnect (InfiniBand or Ethernet). 

Take Networking Courses

Explore deep technical training topics on NVIDIA Quantum  InfiniBand networking through the NVIDIA Academy.

Ready to Purchase?

NVIDIA LinkX products provide AI-optimized connectivity for NVIDIA Spectrum-X switches, Quantum switches, ConnectX NICs, and BlueField™ DPUs.