Accelerated by the demand for large-scale AI model training and High-Performance Computing (HPC), data center networks are rapidly evolving toward 800G and 1.6T speeds. With NVIDIA ConnectX-8 (CX-8) NICs delivering throughput up to 800Gbps per card, AICPLIGHT provides a comprehensive suite of optical modules, Active/Passive Copper Cables, and fiber patch cords to ensure seamless high-speed connectivity.
ConnectX-8 NIC Specification
ConnectX-8 NIC is available in three primary configurations tailored for different deployment densities:
| NIC Model |
Interface Type |
Max Throughput |
Recommended Solution |
| C8180 |
1 × OSFP |
800Gbps |
OSFP (224G SerDes) |
| C8240 |
2 × QSFP112 |
800Gbps |
QSFP112 (112G SerDes) |
| C8220 |
2 × QSFP112 |
400Gbps |
QSFP112 (112G SerDes) |
Among them, the C8180 is the key building block for next-generation InfiniBand XDR (800G) deployments and 1.6T scaling architectures.
C8180 Interconnection: 1 x OSFP Cage (800G XDR or 2 x 400GbE)
The C8180 NIC provides a single OSFP cage, supporting up to:
1.6T InfiniBand XDR Architecture: 1.6T Switch Port Interconnected with 2 x C8180 NICs (XDR 800G)
Taking the NVIDIA Quantum-X800 Q3400-RA (Quantum-3 based) switch as the primary example, this platform features 144 XDR ports distributed over 72 OSFP cages. By leveraging a 1-to-2 splitting strategy, each OSFP cage achieves a 1.6T aggregate rate via two logical XDR 800G ports. This configuration seamlessly interconnects with two independent ConnectX-8 C8180 SuperNICs at 800Gb/s per link to deliver full 1.6T data throughput.
Figure 1: Interconnection diagram showing an NVIDIA Q3400-RA 1.6T switch port splitting into two 800G links for ConnectX-8 C8180 SuperNICs using AICPLIGHT OSFP224 transceivers and MPO-12 fiber cables.
Optical Interconnect Solution:
For high-performance inter-rack connectivity requiring low loss and anti-interference capabilities, the following optical components are recommended:
-
Switch Side: Uses a 1.6T 2x 800Gb/s Twin-port OSFP224 optical module (2 x XDR 800G) finned top. Quantum-X800 Q3400-RA switch can be equipped with AICPLIGHT OSFP-1.6T-2DR4 OSFP224 InfiniBand XDR optical transceiver. This module utilizes 200G-PAM4 modulation and features a Finned Top design for optimized thermal management in high-density switch environments.
-
NIC Side: Uses an OSFP224 optical module with a single-fiber interface (XDR 800G) flat top. Each of the two C8180 NICs utilizes an AICPLIGHT OSFP-800G-DR4 OSFP224 InfiniBand XDR optical transceiver. These modules feature a Flat Top design to fit the standard NIC form factor.
-
Optical Cabling: Connectivity is established via two single-mode OS2 MPO-12/APC fiber patch cords (Type B).
Copper Interconnect Solution:
For intra-rack deployments where power efficiency and cost-effectiveness are prioritized, the following high-speed copper alternatives can be used:
-
DAC Solution (≤1m): 1.6T OSFP to 2x800G OSFP Direct Attach Copper (DAC) cable, providing the lowest latency and power consumption for short-reach connections.
-
AEC Solution (≤3m): 1.6T OSFP to 2x800G OSFP Active Electrical Cable (AEC), offering superior cable flexibility and thinner diameters to improve airflow management.
1.6T InfiniBand NDR Architecture: 1.6T Switch Port Interconnected with 4 x C8180 NICs (NDR 400G)
Utilizing the NVIDIA Quantum-X800 Q3200-RA (Quantum-3 based) switch as a high-density example, this platform supports 36 XDR ports via 18 OSFP cages. By implementing a 1-to-4 splitting strategy, a single 1.6T OSFP cage can interconnect with four independent ConnectX-8 C8180 SuperNICs. This configuration enables an aggregate throughput of 1.6T while providing higher port density by operating each link at NDR 400G speed.
Figure 2: Technical diagram of a 1.6T InfiniBand NDR architecture interconnecting an NVIDIA Quantum-X800 Q3200-RA switch port with four ConnectX-8 C8180 SuperNICs using AICPLIGHT 1.6T 2xDR4 and 800G DR4 (configured for 400G) OSFP224 transceivers via MPO-12 breakout cables.
Optical Interconnect Solution:
For high-density inter-rack deployments requiring extended reach and signal integrity, the following optical components are recommended:
-
Switch Side: Uses a 1.6T 2x 800Gb/s Twin-port OSFP224 optical module (2 x XDR 800G) finned top. Quantum-X800 Q3200-RA switch employs AICPLIGHT OSFP-1.6T-2DR4 OSFP224 InfiniBand XDR optical transceiver. This module features a Twin-port design with 200G-PAM4 modulation and a Finned Top for advanced heat dissipation.
-
NIC Side: Uses a single-fiber interface OSFP module flat top. Each of the four C8180 NICs utilizes an AICPLIGHT OSFP-800G-DR4 OSFP224 optical transceiver. To match the architecture, these Flat Top modules must be configured to operate at 400G speed.
-
Optical Cabling: Connectivity is facilitated through MPO-12/APC Y-splitter (1-to-2) fiber patch cords.
Copper Interconnect Solution:
For cost-effective, short-reach connectivity within a single rack, robust copper alternatives can be used:
-
DAC Solution (≤1m): 1.6T OSFP to 4 x 400G OSFP Direct Attach Copper (DAC), ideal for localized server-to-switch links with zero power consumption from the cable itself.
-
AEC Solution (≤3m): 1.6T OSFP to 4 x 400G OSFP Active Electrical Cable (AEC), designed for easier routing and improved cable management in dense AI clusters.
800G Network Architecture: 800G Switch Port Interconnected with 2 x C8180 NICs (2 x 400G)
Utilizing the NVIDIA SN5610 Ethernet switch as a benchmark, this platform delivers 64 ports of 800GbE in a dense 2U form factor, fully splittable to up to 128 ports of 400GbE, and up to 256 ports of 50/100/200GbE when used with splitter cables.
By implementing a 1-to-2 splitting strategy, a single 800G OSFP port can be fully split to support two independent ConnectX-8 C8180 SuperNICs. This configuration allows for flexible scaling, supporting up to 128 ports of 400GbE across the entire switch fabric.
Figure 3: Technical diagram showing an 800G Ethernet architecture interconnecting an NVIDIA Spectrum-4 SN5610 switch port with two ConnectX-8 C8180 SuperNICs. The solution features AICPLIGHT 800G 2xDR4 (Finned Top) and 400G DR4 (Flat Top) OSFP transceivers connected via MPO-12/APC single-mode fiber patch cords.
Optical Interconnect Solution:
For reliable inter-rack connectivity with high signal integrity and anti-interference capabilities, the following optical components are recommended:
-
Switch Side: Employs AICPLIGHT OSFP-800G-2DR4 OSFP Finned Top optical transceiver. This module utilizes a dual-fiber interface (2 x 400G) to link directly to the SN5610 switch ports.
-
NIC Side: Each of the two C8180 NICs utilizes an AICPLIGHT OSFP-400G-DR4 OSFP Flat Top optical transceiver. These modules provide a single-fiber interface designed for high-density NIC deployments.
-
Optical Cabling: Connectivity is established via two single-mode OS2 MPO-12/APC fiber patch cords (Type B).
Copper Interconnect Solution:
For cost-effective, high-efficiency intra-rack or server-to-ToR connections, the following high-performance copper alternatives can be utilized:
-
DAC Solution (≤2m): 800G OSFP to 2 x 400G OSFP Direct Attach Copper (DAC), providing a stable and efficient link with zero cable power consumption.
-
AEC Solution (≤7m): 800G OSFP to 2 x 400G OSFP Active Electrical Cable (AEC), offering extended reach and superior cable flexibility for complex airflow management.
C8240 Interconnection: 2 x QSFP112 Cages (2 x 400G)
The C8240 NIC provides two QSFP112 cages, each supporting 400G. When both ports are used simultaneously, it supports up to 800G.
800G Network Architecture: 800G Switch Port Interconnected with C8240 NIC (2 x 400GbE / 2 x NDR 400G)
In this configuration, the NVIDIA Spectrum-4 SN5610 switch utilizes its high-density OSFP ports to support the ConnectX-8 C8240 series. By implementing a 1-to-2 splitting strategy, a single 800G OSFP port on the switch is divided into two 400G links, each connecting to one of the dual QSFP112 ports on the C8240 NICs. This setup is ideal for optimizing bandwidth distribution across multi-port NIC environments.
Figure 4: Technical diagram illustrating an 800G Ethernet architecture interconnecting an NVIDIA Spectrum-4 SN5610 switch port with a ConnectX-8 C8240 NIC. The setup features an AICPLIGHT 800G 2xDR4 (Finned Top) OSFP transceiver and two 400G DR4 (Flat Top) QSFP112 transceivers connected via MPO-12/APC fiber cables.
Optical Interconnect Solution:
For high-performance inter-rack connectivity requiring robust signal integrity over longer distances, the following optical components are recommended:
-
Switch Side: SN5610 switch employs AICPLIGHT OSFP-800G-2DR4 OSFP Finned Top optical transceiver. This module provides a dual-fiber interface (2 x 400G) specifically designed for breakout applications.
-
NIC Side: Each of the two C8240 NIC ports utilizes an AICPLIGHT Q112-400G-DR4 QSFP112 Flat Top optical transceiver. These modules are purpose-built for the QSFP112 form factor to deliver dedicated 400G throughput.
-
Optical Cabling: Connectivity is established via two single-mode OS2 MPO-12/APC fiber patch cords (Type B).
Copper Interconnect Solution:
For high-efficiency, short-reach intra-rack connectivity, AICPLIGHT offers specialized breakout copper solutions:
-
DAC Solution (≤2m): 800G OSFP to 2 x 400G QSFP112 Direct Attach Copper (DAC), ensuring ultra-low latency and zero power consumption for rack-level integration.
-
AEC Solution (≤7m): 800G OSFP to 2 x 400G QSFP112 Active Electrical Cable (AEC), providing extended reach and thinner cable diameters for superior cable management in dense AI racks.
C8220 Interconnection: 2 x QSFP112 Cages (400G or 2 x 200G)
The NVIDIA ConnectX-8 C8220 NIC is engineered for next-generation AI and HPC workloads, providing dual QSFP112 ports that support flexible bandwidth configurations: either 1 x 400G (Single Port) or 2 x 200G (Dual Ports).
800G Network Architecture: 800G Switch Port Interconnected with 2 x C8220 NICs (400GbE / NDR 400G)
In high-density AI clusters using the NVIDIA Spectrum-4 SN5610 switch, a single 800G OSFP port can be effectively split into two 400G links to connect multiple C8220 SuperNICs. This 1-to-2 breakout architecture maximizes port efficiency while maintaining low-latency 400GbE or NDR 400G connectivity.
Figure 5: High-performance AI networking interconnect diagram - NVIDIA SN5610 800GbE OSFP switch connected to two ConnectX-8 C8220 400Gb/s SuperNICs using 800G 2xDR4 OSFP to 400G DR4 QSFP112 breakout optical solution with MPO-12 SMF cables.
Optical Interconnect Solution:
-
Switch Side: Utilizes an OSFP-800G-2DR4 optical transceiver (Finned Top). This module features dual MPO-12 interfaces, each delivering 400G throughput via 4-lane 112G SerDes technology.
-
NIC Side: Employs two 400G DR4 QSFP112 optical transceivers (Flat Top). These modules are specifically designed for the QSFP112 cages found on C8220 NICs, ensuring optimal thermal management.
-
Optical Cabling: Connected via two OS2 MPO-12/APC single-mode fiber patch cords, supporting reliable transmission over extended distances.
Copper Interconnect Solution:
For intra-rack or cross-rack connections where power efficiency and latency are critical, the following copper interconnects are recommended:
-
DAC Solution (≤2m): 800G OSFP to 2 x 400G QSFP112 DAC, the ideal zero-power solution for direct-attach connections within the same rack.
-
AEC Solution (≤7m): 800G OSFP to 2 x 400G QSFP112 AEC features active retiming chips to overcome signal integrity challenges of 112G SerDes, allowing for longer reaches across adjacent racks.
800G Network Architecture: 800G Switch Port Interconnected with 2 x C8220 NICs (2 x 200GbE or 2 x NDR 200G Mode)
In scenarios requiring high-density distribution at lower per-port speeds, the NVIDIA Spectrum-4 SN5610 switch can be configured to support the ConnectX-8 C8220 series. By implementing a 1-to-4 splitting strategy, a single 800G OSFP switch port breaks out into four 200G lanes, providing balanced bandwidth across the dual QSFP112 ports of two C8220 NICs. This architecture is optimized for large-scale GPU clusters where port count and efficiency are paramount.
Figure 6: Technical diagram of an 800G Ethernet architecture interconnecting an NVIDIA Spectrum-4 SN5610 switch with two ConnectX-8 C8220 NICs. The setup uses a 1-to-4 breakout strategy, splitting an 800G OSFP Finned Top transceiver into four 200G lanes across four QSFP112 Flat Top transceivers via MPO-12 Y-splitter fiber cables.
Optical Interconnect Solution:
For high-density inter-rack connectivity with optimized signal distribution, the following optical components are recommended:
-
Switch Side: Employs an AICPLIGHT OSFP-800G-2DR4 OSFP Finned Top optical transceiver. This module utilizes a dual-fiber interface (2 x 400G) to facilitate the high-speed breakout transition.
-
NIC Side: Each of the four NIC ports (across two C8220 cards) utilizes an AICPLIGHT Q112-400G-DR4 QSFP112 Flat Top optical transceiver. These modules are configured to operate at 200G per port to match the split bandwidth requirement.
-
Optical Cabling: Connectivity is established via MPO-12/APC Y-splitter (1-to-2) fiber patch cords.
Copper Interconnect Solution:
For high-efficiency, short-reach intra-rack connectivity, AICPLIGHT provides specialized multi-port breakout copper alternatives:
-
DAC Solution (≤2m): 800G OSFP to 4 x 200G QSFP112 Direct Attach Copper (DAC), offering the most cost-effective and low-latency path for rack-level scaling.
-
AEC Solution (≤7m): 800G OSFP to 4 x 200G QSFP112 Active Electrical Cable (AEC), featuring active signal conditioning to ensure reliable 200G data integrity over longer cable lengths.
Summary
When deploying multiple CX-8 NICs per node, aggregate bandwidth can scale to 800G and 1.6T to meet AI/HPC workload demands. In actual deployment, choose between optical or copper solutions based on transmission distance and performance requirements. Generally, copper solutions are used for intra-rack interconnections, while optical modules and fiber patch cords are preferred for inter-rack connectivity.
English
