GS-NET-04Engineering datasheet

NVIDIA ConnectX-7

Statusshipping
Verified2026-07-10

ConnectX-7 is the node-level network adapter of the H100/H200 generation: a 400 Gb/s PCIe Gen5 card that runs InfiniBand NDR or Ethernet from the same silicon (VPI). In a rail-optimized HGX design, one ConnectX-7 sits behind each GPU, PCIe-switch-attached so GPUDirect RDMA traffic never crosses the CPU root complex. The connector and protocol you order lock in the fabric you're building toward, InfiniBand for maximum determinism or Spectrum-X Ethernet for operational uniformity, so it is worth getting that decision right before ordering eight of these per node.

400 Gb/s
Max bandwidth · single-port NDR InfiniBand or 400GbE
PCIe 5.0 x16
Host interface · Gen4 x16 compatible
330-370 Mmsg/s
Message rate · reported, RDMA
1, 2, or 4
Ports · OSFP, QSFP112, or SFP56 depending on SKU
~25 W
Typical power · single 400G port, passive cables
Interfaces01/05
Host interfacePCIe Gen 5.0, x16 · 16/32 GT/s, Gen 3.0 compatible
InfiniBand speedsNDR 400, HDR 200, EDR 100 Gb/s
Ethernet speeds400/200/100/50/25/10 GbE
ConnectorsOSFP, QSFP112, or SFP56 · SKU-dependent, not interchangeable
Interface technologiesNRZ (10/25G), PAM4 (50/100G)
Performance & offloads02/05
RDMARoCEv2, zero-touch RoCE
GPU offloadGPUDirect RDMA, GPUDirect Storage
Message rate330-370 Mmsg/s · reported
Collective offloadSHARP · in-network reduction
SDN offloadASAP2 · OVS/VXLAN/GENEVE acceleration
Security offloadIPsec, TLS, MACsec · 128/256-bit, in-line
Form factors03/05
PCIe cardsHHHL, FHHL
OCP cardsOCP 3.0 TSFF, SFF
Multi-HostUp to 4 hosts · via PCIe bifurcation, 32-lane variants
Power04/05
Single-port 400G, typical24.9-25.9 W · OSFP or QSFP112, passive cables
Dual-port variants, typical24.9-25.9 W · MCX755106 series
Operating temperature0-55°C
Storage temperature-40-70°C
Platform support05/05
Reference deployment1 NIC per GPU · rail-optimized HGX H100/H200 nodes
OS supportLinux (inbox/MLNX_OFED), Windows, VMware ESXi
Fabric compatibilityInfiniBand NDR fabrics or Spectrum-X Ethernet
Field notes
  • NIC-per-GPU cost math is easy to leave out of a cluster estimate: 8× ConnectX-7 at $1,800-$3,700 street adds $14K-$30K per 8-GPU node before switch ports, optics, and cabling are counted.
  • InfiniBand vs. Spectrum-X Ethernet is decided at the fabric level, not the NIC: ConnectX-7 (VPI) runs either. If the far end is an SN5600 Spectrum-X switch, the card's Ethernet firmware and RoCE congestion-control behavior matter more than raw port speed.
  • GPUDirect RDMA needs a PCIe-switch topology, NIC and GPU sharing a PCIe switch downstream of the CPU, to avoid a CPU root-complex hop. That is a server board design decision the NIC alone cannot fix.
  • Connector choice locks in a cabling ecosystem early: single-port OSFP variants target InfiniBand NDR fabrics, dual-port QSFP112 variants target Ethernet-centric or mixed deployments. Changing later means re-cabling, not re-flashing.
  • At roughly 25 W typical per card, power draw is a rounding error next to the GPU it feeds, but 8 cards per node across a few hundred nodes adds a real line item to PDU and airflow planning.
Questions we get on this part

What is ConnectX-7 used for?

It's the per-GPU network adapter in H100/H200-class clusters, typically one card per GPU in a rail-optimized design, running either InfiniBand NDR (400 Gb/s) or 400GbE.

How much does a ConnectX-7 cost?

Reported street pricing is roughly $1,800 to $3,700 per card depending on reseller and configuration; single-port 400GbE QSFP112 cards run near the low end of that range at catalog resellers.

Should I choose InfiniBand or Ethernet for ConnectX-7?

The card supports both from the same silicon (VPI); the decision is made at the fabric level. InfiniBand stays more deterministic for tight GPU collective synchronization. Spectrum-X Ethernet (ConnectX-7 or BlueField-3 paired with SN5600 switches) trades a little of that determinism for standard Ethernet tooling and multi-vendor optics.

Does ConnectX-7 need BlueField-3 to work?

No. ConnectX-7 is the standard host NIC used across both InfiniBand and Spectrum-X deployments. BlueField-3 is a separate DPU-class SuperNIC that NVIDIA's Spectrum-X reference designs use specifically to unlock full adaptive routing and congestion control; ConnectX-7 works with or without it elsewhere in the fabric.