Lenovo 4XC7A99368 Thinksystem NVIDIA CONNECTX-7 NDR OSFP400 1-PORT PCIE GEN5 X16 Infiniband Adapte

Lenovo 4XC7A99368 ConnectX-7 NDR 400Gb/s Single-Port PCIe Gen5 InfiniBand Adapter

Overview

The Lenovo 4XC7A99368 is a single-port OSFP400 InfiniBand adapter built on the ConnectX-7 silicon, delivering NDR (Next Data Rate) 400Gb/s fabric connectivity over a PCIe Gen5 x16 host interface. It is designed for ThinkSystem servers in environments where fabric bandwidth is the primary bottleneck — dense AI training clusters, high-performance computing (HPC) nodes, and scale-out storage architectures that have exhausted what 100GbE or HDR InfiniBand can offer. The OSFP400 port format keeps cabling options broad across NDR switch infrastructure, and the PCIe Gen5 x16 slot assignment ensures the host bus does not become the constraint for a 400Gb/s link.

Key Features

• NDR 400Gb/s InfiniBand: NDR doubles the per-port throughput of HDR (200Gb/s) — directly relevant when GPU-to-GPU or storage-to-compute traffic regularly saturates your existing fabric. A single port delivering 400Gb/s means fewer adapters per node without sacrificing bandwidth headroom.
• PCIe Gen5 x16 Host Interface: PCIe Gen5 x16 provides approximately 128 GB/s of theoretical host bandwidth, keeping the CPU-to-adapter path from throttling the 400Gb/s wire speed. In prior-generation HDR deployments on Gen4 hosts, the PCIe lane often became the ceiling — Gen5 removes that constraint.
• OSFP400 Port Format: The OSFP (Octal Small Form-factor Pluggable) connector supports both direct-attach copper (DAC) and active optical cable (AOC) assemblies at 400Gb/s, giving you cabling flexibility across short rack-to-rack and longer inter-rack runs without changing the adapter.
• Single-Port Design: One OSFP400 port per card keeps the slot-to-bandwidth ratio high — useful when PCIe slot count per server is limited and you need maximum throughput from each allocated slot rather than spreading bandwidth across two slower ports.
• ThinkSystem Platform Integration: As a Lenovo-sourced option, the 4XC7A99368 is validated for ThinkSystem server configurations, which simplifies firmware coordination and support stack alignment compared to third-party adapter sourcing. For broader context, explore the Lenovo server and networking catalog or the network adapter category.

Integration and Compatibility

The PCIe Gen5 x16 interface requires a host system with a Gen5-capable slot — Gen4 systems will accept the card electrically but will operate at Gen4 speeds, cutting available host bandwidth roughly in half. NDR fabric requires NDR-capable switches and cables; this adapter does not interoperate at NDR rates with HDR switches, though InfiniBand's backward compatibility allows the link to negotiate to a lower rate if needed. For planning a complete high-speed fabric deployment, reviewing a high-speed fabric planning guide alongside switch and cable selection is strongly advised. Pair with compatible NDR InfiniBand cables and transceivers for a complete validated stack.

Frequently Asked Questions

Q: What InfiniBand data rate does the 4XC7A99368 support?

A: The 4XC7A99368 supports NDR (Next Data Rate) at 400Gb/s per port. This is the current top InfiniBand speed tier, doubling the 200Gb/s of the prior HDR generation.

Q: What PCIe slot does the 4XC7A99368 require?

A: It requires a PCIe Gen5 x16 slot. The card will seat in a Gen4 x16 slot but will operate at Gen4 bandwidth, which reduces the host-side throughput ceiling.

Q: How many InfiniBand ports does the 4XC7A99368 have?

A: One OSFP400 port. The single-port design maximizes per-slot bandwidth rather than splitting throughput across two ports.

Q: What connector type does the 4XC7A99368 use?

A: OSFP400 (Octal Small Form-factor Pluggable). This connector is compatible with NDR DAC and AOC cables from major InfiniBand cable vendors.

Q: Is the 4XC7A99368 compatible with non-Lenovo servers?

A: The card uses a standard PCIe Gen5 x16 interface. While Lenovo validates it for ThinkSystem platforms, physical PCIe compatibility with other server vendors is possible, but firmware and support are optimized for ThinkSystem configurations. Verify compatibility with your specific server vendor before deploying in non-ThinkSystem systems.

Karl Wilson

The 4XC7A99368 lands at the intersection of two simultaneous generational jumps — NDR 400Gb/s InfiniBand and PCIe Gen5 x16 — which is what makes it worth the attention of anyone building or expanding a GPU cluster today rather than 18 months from now. Most HDR deployments I've seen are already feeling the fabric ceiling with 8-GPU nodes; this card doubles the per-port headroom without doubling your slot count.

Technical Highlights:

• NDR 400Gb/s throughput: Per-port bandwidth doubles compared to HDR 200Gb/s — for all-reduce operations in distributed training, that gap compounds across hundreds of nodes and can meaningfully cut synchronization wait time.
• PCIe Gen5 x16 host interface: Roughly 128 GB/s theoretical host bandwidth means the slot itself isn't your bottleneck at 400Gb/s wire speed — a limitation that caught a number of HDR deployments on Gen4 hosts off guard.
• Single OSFP400 port: One port per card keeps the physical slot budget lean; in 2U servers with limited PCIe real estate, that matters more than it sounds when you're populating eight GPU slots plus NVMe controllers in the same chassis.

Deployment Considerations:

• Confirm your ThinkSystem model has a PCIe Gen5 x16 open slot before ordering — not all current ThinkSystem configurations expose Gen5 slots, particularly in entry-tier 1U platforms.
• NDR switch infrastructure is a hard requirement for NDR link speed; dropping this into an existing HDR fabric will negotiate down to HDR rates, which eliminates the bandwidth advantage and makes the investment premature.

The 4XC7A99368 is the right call for new ThinkSystem-based AI training builds where the fabric is being sized for 400Gb/s from the start — not for HDR refresh projects where the switch infrastructure isn't also being upgraded.