NVIDIA ConnectX-8 SuperNIC: specs, 800 Gb/s ports, PCIe Gen6 and its job in an AI server

What the ConnectX-8 SuperNIC is

The NVIDIA ConnectX-8 SuperNIC is a network adapter for AI servers that runs InfiniBand or Ethernet at up to 800 Gb/s and connects to the host over PCIe Gen6. At full rate it offers one 800 Gb/s XDR InfiniBand port or, as Ethernet, two 400GbE links on one OSFP cage. In NVIDIA's Blackwell systems it sits next to the GPUs: eight in a DGX B300, eight on the HGX B300 baseboard, and in the 4U and 6U RTX PRO Server designs the GPUs connect through its PCIe Gen6 switch.

SuperNIC is NVIDIA's name for the class: its product page (page date 24 November 2025) calls it "a new class of network accelerators designed to supercharge hyperscale AI workloads". NVIDIA's blog of 15 October 2024 names what such an adapter adds for traffic between GPU servers: RoCE adaptive routing, direct data placement, congestion control with switch telemetry and inline encryption with IPsec and PSP at up to 800 Gb/s.

ConnectX-8 SuperNIC specs in one table

NVIDIA ConnectX-8 SuperNIC datasheet (document 4248150, September 2025); NVIDIA ConnectX-8 SuperNIC User Manual, pages Specifications and Introduction (30 September 2026) and Supported Interfaces (30 October 2026); read 12 November 2026.

The two documents differ on encryption. The datasheet lists MACsec next to IPsec and PSP, while the manual's feature text names IPsec and PSP acceleration and mentions MACsec only as the standard IEEE 802.1AE. Inline TLS appears in neither.

800G on the ConnectX-8: one XDR port or two 400GbE

The 800 Gb/s in the datasheet are the total of the card. As InfiniBand, that is a single XDR port. As Ethernet, the same OSFP cage carries two 400GbE links, four at 200GbE or eight at 100GbE. The manual's list of Ethernet rates ends at 400 GbE, and neither the port table nor the datasheet names an 800GbE port, so a tender should ask for 2 × 400GbE.

NVIDIA ConnectX-8 SuperNIC User Manual, Port Configurations (11 November 2026), read 12 November 2026.

NVIDIA calls these configurations "tested and verified" and refers customers with unlisted configurations to NVIDIA Support. The XDR Lite modes at 400 Gb/s are, by the same page, for hosts up to PCIe Gen 5.0.

PCIe Gen6 x16 on the card, 48 lanes on the chip

Two lane counts are in circulation for the ConnectX-8. The add-in cards have an x16 edge connector that runs PCIe Gen6 at 64 GT/s and is compatible with Gen5 and Gen4 hosts. The datasheet's "PCIe Gen6 (up to 48 lanes)" is the figure for the chip; it matters in switch-board and baseboard designs, where the GPUs hang off the ConnectX-8.

The manual's page on bifurcation describes two x16 interfaces "with an internal PCIe switch architecture": the edge connector, by default an endpoint to the host, and an MCIO connector. NVIDIA documents two uses for the second one.

• Socket Direct. The 32-lane bus is split into two x16 links for the two processors of a dual-socket server. By the manual, this supports 800 Gb/s in PCIe Gen5 servers and 400 Gb/s in Gen4 servers. It needs the auxiliary kit 930-9XAX6-0025-000 with a 250 mm MCIO harness, which is not in the card's package.
• Downstream port. The MCIO connector becomes a root complex for storage devices, by default for four SSDs on PCIe Gen5 at 32 GT/s, and the SuperNIC works as a PCIe switch with NVMe drives as endpoints.

A Gen5 server needs both links for the full rate. A PCIe Gen5 x16 link carries about 64 GB/s in each direction, the figure our guide to CPU, RAM and storage for a 4-GPU server works with; that is about 512 Gb/s by our arithmetic, less than the 800 Gb/s of the port. The bifurcation page lists one x16 link at Gen6 as the default and two at Gen5 as the recommended option, and warns: "Misconfiguration may harm the system."

Network cards are part of the servers we build to order. Send us the port speed, the protocol and the PCIe generation of the host, and our team replies within one business day with a configuration that names the adapter and its part number.

The PCIe switch for GPUs in RTX PRO Servers

NVIDIA's own emphasis is on a different switch role, with the GPUs behind the SuperNIC. Its developer blog of 18 May 2025 describes the ConnectX-8 as "integrating 48 lanes of PCIe Gen6 connectivity with an integrated PCIe Gen6 switch". The comparison is a server with two CPUs, eight GPUs and five network adapters, a layout that "requires two to four discrete PCIe switches". With the GPUs behind ConnectX-8 SuperNICs, the blog gives up to 400 Gb/s of network bandwidth per GPU at a GPU-to-NIC ratio of 2:1. Our articles on the NVIDIA RTX PRO Server and on NVIDIA MGX servers compare the two layouts, and one model on several GPUs shows which workloads send data between cards.

Two points keep a specification accurate. The NVIDIA documents we read state no number of ConnectX-8 SuperNICs for the RTX PRO Server switch board. And the claim that the ConnectX-8 is the first network adapter with a PCIe switch has no support in them: the ConnectX-7 datasheet already lists "Integrated PCI switch", and what NVIDIA describes as new is Gen6, 48 lanes and GPUs behind the SuperNIC.

ConnectX-8 vs ConnectX-7

NVIDIA ConnectX-7 datasheet (marked APR21/MAY21) and ConnectX-7 Adapter Cards User Manual, Introduction and Specifications (12 August 2026); ConnectX-8 datasheet (September 2025) and User Manual; read 12 November 2026.

Against the ConnectX-7, the ConnectX-8 doubles the total bandwidth and the PCIe signalling rate and adds XDR InfiniBand. Two rows run the other way: Multi-Host goes from eight hosts to four, and inline TLS is no longer listed, while PSP is new. The ConnectX-7 power figure is the manual's typical value for the single-port OSFP card with passive cables on PCIe Gen5 x16; its crypto variant MCX75310AAC-NEAT is listed with 25.9 W.

ConnectX-8 models, form factors and part numbers

The datasheet names four form factors: "PCIe HHHL 1P x OSFP", "PCIe HHHL 2P x QSFP112", "Dual ConnectX-8 Mezzanine" and "OCP 3.0 TSFF 1P x OSFP". The user manuals list the following cards with NVIDIA's part numbers.

NVIDIA ConnectX-8 SuperNIC User Manual, index (4 November 2026) and Specifications (30 September 2026); ConnectX-8 SuperNIC for OCP 3.0 User Manual, index (14 October 2026) and Specifications (9 February 2026); read 12 November 2026.

The PCIe cards are half height and half length, and all listed part numbers carry crypto and Secure Boot. The two OCP 3.0 cards measure 115.00 × 76.00 mm; their manual marks them "Prototype" in the lifecycle column and gives their power as "TBD", so we do not treat them as a part to order.

The manual's pages disagree on one point. The part-number table on the index page assigns the downstream-port option to 900-9X81E-00EX-DT0 and Socket Direct to the C8240, while the Introduction links the downstream port to the C8240 and C8220 and Socket Direct and Multi-Host to the C8180 and C8180L. We follow the part-number table; if a design depends on one of the options, have the part number confirmed in the quote. NVIDIA's 900-codes on GPUs are the subject of our article on NVIDIA GPU part numbers.

Where NVIDIA systems use the ConnectX-8

NVIDIA DGX B300 User Guide, introduction; NVIDIA HGX AI Factory enterprise reference architecture, Components and appendix (18 May 2026); NVIDIA RTX PRO Servers product page (11 June 2026); Gigabyte product page XL44-SX2-AAS1 (© 2026); NVIDIA DGX Station datasheet 5543600 (August 2026); read 12 November 2026.

The DGX B300 user guide states: "ConnectX-8 networking connects directly to GPUs via PCIe Gen6." Its port table shows two ports per OSFP cage; the rest of that system is in our DGX B300 specifications. In the HGX B300 the SuperNICs are part of the board, so a server built on it, such as the Supermicro SYS-822GS-NB3RT, brings its east-west ports with it.

For RTX PRO Servers, two NVIDIA documents describe different networks. The product page of 11 June 2026 lists the MGX 4U and 6U designs with ConnectX-8 SuperNICs and a BlueField-3 DPU, and the two MGX 2U designs with ConnectX-7: one with two RTX PRO 6000 Blackwell Server Edition cards and ConnectX-7 or BlueField-3 SuperNICs, one with eight RTX PRO 4500 Blackwell Server Edition cards (its specifications and our page to buy the RTX PRO 4500 Server Edition). The RTX PRO AI Factory reference architecture of 18 May 2026 still specifies four BlueField-3 SuperNICs and one BlueField-3 DPU per server and notes: "With Gen 6 PCIe switches, higher network bandwidth with up to 400Gbps per GPU can be supported." What the DPU is for is the subject of our article on the BlueField-3 DPU in GPU servers.

Gigabyte's page for the XL44-SX2-AAS1 lists eight RTX PRO 6000 Blackwell Server Edition GPUs "via ConnectX®-8"; our overview of Gigabyte servers for the RTX PRO 6000 Server Edition covers the range. For the DGX Station, no NVIDIA document we read says that the ConnectX-8 acts as a PCIe switch; the two ports link two systems, as which models run on the DGX Station GB300 describes.

RoCE and Spectrum-X features by name

For RDMA the datasheet lists RoCEv2 acceleration, GPUDirect RDMA and GPUDirect Storage, programmable congestion control based on telemetry, high-speed packet reordering and, on InfiniBand models, SHARP. The libraries it names are NCCL, HPC-X, DOCA UCC/UCX, Open MPI and MVAPICH2.

Spectrum-X is NVIDIA's Ethernet platform of switches and SuperNICs. NVIDIA's Spectrum-X page of 22 October 2026 names a SuperNIC per GPU generation, BlueField-3 for Hopper, ConnectX-8 for Blackwell and ConnectX-9 for Vera Rubin NVL72. It defines adaptive routing as "a feature where the Spectrum-X Ethernet switch and SuperNIC work in tight coordination", so this feature needs both ends, and it names "advanced telemetry-based congestion control to provide noise isolation between tenants".

One limit comes from NVIDIA's DOCA documentation of 1 September 2026: ConnectX devices cannot use the DPA libraries, the programmable data path that the BlueField-3 offers.

Power, temperature and where the card may be installed

NVIDIA publishes no wattage for the ConnectX-8 in the public manual or the datasheet. The manual refers to the "NVIDIA ConnectX-8 SuperNIC Product Specifications", which NVIDIA provides through NVOnline or a representative. The public documents give the supply voltages and an operating temperature of 0 to 55 °C. For the power and airflow budget of a slot, ask the server maker for the figure of the exact card; we print none. The manual also limits where the card goes: it is "designed and validated for operation in data-center servers" and not intended for desktops or workstations.

Firmware, drivers and operating systems

NVIDIA's firmware release notes of 31 August 2026 list 40.50.1002 as the newest ConnectX-8 firmware and 40.47.3573 as the long-term line ("2025 LTS U3"). The host software paired with it is DOCA-Host 3.5.0 and, for Windows, WinOF-2 26.7.50000; the table has no MLNX_OFED row. Within DOCA-Host, the profile doca-ofed is the "Driver-only installation equivalent to MLNX_OFED", and doca-networking is the profile recommended for ConnectX adapters.

DOCA 3.5.0 lists as host systems, among others, Ubuntu 22.04, 24.04 and 26.04, RHEL and Rocky Linux 8.10, 9.x and 10.x, SLES 15 SP6, SP7 and 16, and Debian 12 and 13 (General Support page, 2 September 2026). Our guide to GPU server firmware updates describes the update order for a node.

When an inference server needs no ConnectX-8

A single inference server with PCIe cards needs none of this. NVIDIA's configuration guide for NVIDIA-Certified Systems, read 12 November 2026, says: "Single node deployments typically do not require high-speed networking to connect multiple nodes for your AI workload." Two to four servers that each run complete model copies behind a load balancer exchange nothing with each other either; our article on a GPU inference cluster without InfiniBand plans two 25 GbE ports per server for that case.

A fast fabric matters when one model spans servers or when KV cache moves between them. For that the guide asks for a "Minimum 200 Gbps for multi-node inference" and names "Up to 400 Gbps per GPU". NVIDIA's reference architecture for servers with eight RTX PRO 6000 Blackwell Server Edition cards gives "4x 200 Gb/s or 2x 400 Gb/s NICs" as the minimum and "4x 400 Gb/s NICs" as recommended; the one for the HGX B300 starts at "8x 400 Gb/s Ethernet NICs". ConnectX-7 cards and BlueField-3 SuperNICs in ordinary PCIe slots also reach 200 or 400 Gb/s per adapter.

The ConnectX-8 enters a plan in three cases: an XDR InfiniBand fabric at 800 Gb/s, an HGX B300 server and an RTX PRO Server with the ConnectX-8 switch board. In the last two the decision is made with the chassis, because the east-west ports are part of the board. In a server with a classic PCIe layout, adapters can be added while the chassis keeps free x16 slots, and the guide asks that NICs sit "within the same PCIe switch or root complex as the GPUs".

We supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project; the switches of the fabric are outside what we supply. Tell us how many servers you plan and which fabric your switches provide, and we answer with the adapters and ports for each server.

FAQ