800G Transceiver Compatibility for NVIDIA Platforms: Spectrum-4, Quantum-X800, ConnectX-8, and BlueField-3

1. Why Compatibility Errors Are Costly

A single mismatched module can stall an entire deployment. NVIDIA's current platform lineup uses different OSFP cage types, different form factors, and different protocols across switches, NICs, and DPUs. A twin-port IHS module that works perfectly in a Quantum-X800 switch will not physically fit in a ConnectX-8 SuperNIC. A QSFP112 module designed for BlueField-3 has nothing to do with the OSFP slots on a Spectrum-4.

2-tier leaf-spine AI cluster architecture: GPU servers with ConnectX-8 NICs connect to Spectrum-4 leaf switches, which uplink to Quantum-X800 spine switches. Each tier uses different OSFP form factors and interconnect types.

2. AI Cluster Architecture: Where Each Platform Sits

In a typical AI data center, GPU servers connect to the network through a structured leaf-spine fabric. Each tier of this fabric uses different NVIDIA platforms, different OSFP form factors, and different interconnect types. Understanding this architecture is the foundation for getting your transceiver orders right.

2-Tier Leaf-Spine (Most Common)

The standard architecture for AI/GPU clusters up to approximately 4,000 GPUs. Three tiers of hardware:

• GPU/Compute Tier: GPU servers with ConnectX-8 SuperNIC (800G XDR, OSFP flat-top). The NIC connects to the leaf switch via DAC (2m or less), LACC (3-5m), AOC (30-100m), or optical transceivers (SR8 100m, DR4 500m).
• Leaf Tier (Top of Rack): Spectrum-4 SN5600 switches (51.2 Tb/s, 64 OSFP). These accept ALL interconnect types — both IHS and RHS OSFP. Breakout capability: each 800G port can split into 2x400G, 4x200G, or 8x100G.
• Spine Tier: For InfiniBand fabrics, Q3400-RA (115.2 Tb/s, 72 twin-port OSFP, IHS only) or Q3200-RA (2x28.8 Tb/s, NDR backward compatible). Leaf-to-spine links use DAC, LACC, AEC (5-7m, switch-to-switch only, approximately 27W per end), or optical transceivers (2xDR4 500m, 2xFR4 2km, 2xLR4 10km).

3-Tier (For Massive Scale)

When GPU count exceeds approximately 4,000 and the spine layer runs out of ports, a super-spine layer is added above the spines. Hyperscalers building 10,000+ GPU clusters use 3-tier fabrics. Future CPO switches like the SN6800 (409.6 Tb/s, 512x800G) are designed for this super-spine role.

Storage/Infrastructure Tier (Side Branch)

BlueField-3 DPUs connect to leaf switches for storage, infrastructure management, and multi-tenant isolation. BlueField-3 uses QSFP112 (400G NDR) — NOT OSFP. This is a completely separate form factor requiring distinct transceiver SKUs.

Legacy Tier

Existing deployments may include ConnectX-7 NICs (400G NDR) on DGX H100 Cedar-7 mezzanine cards. CX-7 uses OSFP (flat-top) or QSFP112 form factor. CX-7 uses NDR-generation (100G-PAM4 per lane) cables and transceivers, NOT XDR-generation (200G-PAM4 per lane). They are not interchangeable.

Architecture Variations by Application

In short: your workload sets the fabric — InfiniBand spines (Quantum-X800) for training and HPC where GPU-to-GPU latency dominates, and Ethernet spines (Spectrum-4) for inference and multi-tenant cloud, with a super-spine tier added only once you cross ~10,000 GPUs.

Complete NVIDIA 800G platform map: bandwidth specs, port counts, cage types, and protocols for all platforms. Use this as the reference before specifying any transceiver order.

3. Spectrum-4 Ethernet Switches (SN5600)

The Spectrum-4 SN5600: NVIDIA's flagship 800G Ethernet switch with 64 OSFP ports and 51.2 Tb/s capacity. The only NVIDIA platform that accepts both IHS and RHS OSFP modules, making it the most flexible transceiver platform in any AI cluster fabric.

The Spectrum-4 platform is NVIDIA's flagship 800G Ethernet switch, delivering 51.2 Tbps aggregate switching capacity across 64 OSFP ports. It is the most versatile platform in the NVIDIA lineup for 800G transceiver compatibility because it supports both IHS (Integrated Heat Sink) and RHS (Riding Heat Sink) OSFP modules.

This flexibility means Spectrum-4 can accept twin-port modules like the 800G 2xDR4 (which requires IHS cages for thermal dissipation) as well as single-port modules like the 800G SR8 or DR4 in RHS form factor. It also supports full breakout capability: each 800G port can be split into 2x400G, 4x200G, or 8x100G using the appropriate breakout cables and transceiver configurations.

4. Quantum-X800 InfiniBand Switches (Q3400-RA)

The Quantum-X800 Q3400-RA: NVIDIA's top-tier InfiniBand switch with 72 twin-port IHS OSFP cages delivering 144 effective 800G ports at 115.2 Tb/s. IHS form factor only — RHS modules will not physically fit.

The Quantum-X800 Q3400-RA (also referred to simply as Q3400) is NVIDIA's top-tier InfiniBand switch, delivering 115.2 Tb/s aggregate bandwidth using 72 twin-port OSFP cages with four Quantum-3 ASICs in a 4U chassis. This platform uses exclusively IHS OSFP cages designed for the higher thermal load of twin-port transceivers.

Compatible XDR transceivers include the MMS4A00-XM (twin-port 2xDR4/DR8, OSFP224 form factor, 1.6T per module — delivering 2×800G links from one physical cage in IHS/Closed Finned Top design). Each twin-port module provides two independent 800G InfiniBand XDR links from a single OSFP cage — this is how the Q3400-RA achieves 144 effective 800G ports from 72 physical cages. RHS (flat-top) modules will not work in the Q3400-RA's IHS cages. Always specify IHS form factor when ordering for this platform.

5. Quantum-X800 Q3200-RA InfiniBand Switch

The Q3200-RA: a 2U enclosure with two independent Quantum-3 switches, each with 18 twin-port IHS OSFP cages. Unlike the Q3400-RA, the Q3200-RA supports NDR backward compatibility — the key differentiator for mixed-generation deployments.

The Q3200-RA (often referred to simply as Q3200) is the second Quantum-X800 switch model — widely deployed in environments integrating with existing NDR infrastructure. It is a 2U air-cooled enclosure containing two independent Quantum-3 switches.

Each switch inside the Q3200-RA has 18 twin-port OSFP cages, providing 36 effective 800G ports per switch and 72 total per enclosure. Each switch delivers 28.8 Tb/s bandwidth. Like the Q3400-RA, the Q3200-RA uses IHS OSFP cages exclusively. However, a critical distinction separates the two: the Q3200-RA supports backward compatibility to the NDR (4x100G-PAM4) product line — including DAC, LACC, multimode transceivers, AOCs, and single-mode transceivers — while the Q3400-RA does not.

The Q3200-RA also supports a broader range of port speeds: 40, 56, 100, 200, 400, and 800 Gb/s. The Q3400-RA supports only 400G and 800G. One important limitation documented in the NVIDIA Q3200-RA user guide: ports 17, 18, 35, and 36 on the Q3200-RA do not support split mode or LACC. Always verify against the latest NVIDIA documentation before final design.

6. ConnectX-8 SuperNIC

The ConnectX-8 SuperNIC: single OSFP RHS (flat-top) port supporting 800G InfiniBand XDR or 2x400GbE. The RHS cage cannot accommodate twin-port IHS modules — a critical distinction when ordering alongside Quantum-X800 switches that use IHS exclusively.

The ConnectX-8 SuperNIC supports 800 Gb/s InfiniBand XDR or 2x400GbE Ethernet through a single OSFP port (the adapter uses the OSFP224 connector along with dual-port QSFP112 variants depending on model). The NIC's OSFP cage is RHS (flat-top) only. The single-port cage design cannot accommodate twin-port IHS modules due to height constraints.

Compatible modules include the MMS4A20-XM800 (single-port DR4, 800G, 500m over single-mode fiber) and other RHS-compatible OSFP modules. If you are deploying ConnectX-8 alongside Quantum-X800 switches, the switches use IHS modules and the NICs use RHS modules. They are not interchangeable.

7. ConnectX-7 / DGX H100

The DGX H100 uses ConnectX-7 on Cedar-7 mezzanine cards with four twin-port OSFP RHS cages. CX-7 operates at 400G NDR — not 800G XDR. NDR (100G-PAM4 per lane) and XDR (200G-PAM4 per lane) modules are not interchangeable despite sharing the OSFP form factor.

ConnectX-7 remains the most widely installed NIC in existing GPU clusters, and customers deploying 800G equipment frequently operate mixed CX-7/CX-8 environments. ConnectX-7 NICs operate at 400 Gb/s — InfiniBand NDR or Ethernet — and are available in both OSFP (single-port, flat-top/RHS) and QSFP112 form factors.

The DGX H100 uses ConnectX-7 on Cedar-7 mezzanine cards with four twin-port OSFP cages (RHS/flat-top) on the front panel. Twin-port OSFP transceivers with 1:2 fiber splitter cables can connect from 800G switches (Q3400-RA / Spectrum-4) down to ConnectX-7 at 400G, allowing mixed-generation fabrics without replacing the switch tier.

All CX-7 and BlueField-3 devices contain both InfiniBand and Ethernet protocols, activated by whichever switch they connect to. CX-7 uses NDR-generation (100G-PAM4 per lane) cables and transceivers, NOT XDR-generation (200G-PAM4 per lane). They are not interchangeable despite both using OSFP form factor.

8. IHS vs RHS: The Physical Difference

IHS (Integrated Heat Sink): The heatsink is built into the transceiver module itself, creating a taller profile. The module's top surface has fins or a heat spreader. Requires a taller OSFP cage. Used in switches where cage height is not constrained.

RHS (Riding Heat Sink): The module has a flat top. A heatsink attached to the cage lid presses down onto the module from above. Used in space-constrained designs like NICs where the OSFP cage height is limited by the PCIe card form factor. Only single-port modules are available in RHS.

OSFP IHS vs RHS cage type comparison for NVIDIA 800G platforms — switches (Spectrum-4, Q3400-RA, Q3200-RA) use the taller IHS cage; NICs (ConnectX-8) use the space-constrained RHS cage. The two are physically incompatible.

9. BlueField-3 DPU

The BlueField-3 DPU uses QSFP112 form factor exclusively — not OSFP. It is completely separate from the OSFP ecosystem used by all other NVIDIA 800G platforms. Budget for a distinct transceiver SKU for every BlueField-3 DPU in your deployment.

BlueField-3 DPUs use QSFP112 form factor exclusively — not OSFP. This maintains compatibility with the existing 400G NDR InfiniBand ecosystem while supporting the DPU's integrated processing architecture. The QSFP112 cages accept 400G NDR InfiniBand transceivers with flat-top (RHS-style) thermal design.

If your deployment includes BlueField-3 DPUs alongside 800G switches and NICs, you will need a separate category of transceivers for the DPU tier. Do not attempt to source OSFP modules for BlueField-3 — they are physically incompatible.

10. DGX SuperPOD Cabling Requirements

A DGX SuperPOD deployment spans multiple NVIDIA platforms simultaneously: GPU nodes with ConnectX-8 NICs, leaf and spine switches (Spectrum-4 or Quantum-X800), and potentially BlueField-3 DPUs for storage or infrastructure management. Each tier has its own transceiver requirements, creating a multi-form-factor procurement challenge.

The GPU-to-leaf connections (ConnectX-8 to Spectrum-4) use RHS OSFP modules. The leaf-to-spine connections may use either IHS or RHS depending on configuration. The storage/infrastructure tier (BlueField-3) uses QSFP112 modules. Planning the full transceiver bill of materials requires mapping each tier explicitly.

The takeaway for procurement: a single SuperPOD needs three different module families at once — RHS OSFP for GPU-to-leaf, IHS (or RHS) OSFP for leaf-to-spine, and QSFP112 for the BlueField-3 storage tier — so build your bill of materials tier by tier rather than ordering one part number for the whole fabric.

11. 800G Copper and Active Cables

Many 800G deployments use copper and active cables for short intra-rack links where optical transceivers are unnecessary or cost-prohibitive. There are four distinct cable types — each with different reach, power draw, latency characteristics, and platform compatibility. Choosing the wrong type can mean incompatibility with specific platforms or unexpected power budget overruns.

Read this table by distance first, then check the platform columns: DAC and LACC work everywhere for the shortest links, AEC adds reach for switch-to-switch hops but is not supported on the ConnectX-8 NIC, and AOC covers the longer 30-100m runs on Spectrum-4 and CX-8 but not on the Quantum-X800 switches.

For a complete guide to selecting between optical transceivers, DAC, LACC, AEC, and AOC for your 800G deployment, see our 800G Data Center Interconnect Selection Guide.

800G optics and cable selection by application: reach, power, fiber type, and platform compatibility for every interconnect type. Critical warnings: Q3400-RA / Q3200-RA IHS-only, ConnectX-8 flat-top only, BlueField-3 QSFP112 only, AEC switch-to-switch only.

12. Full Compatibility Matrix

Note: The 2xDR4 and DR8 designations shown below represent two operating modes of the same twin-port OSFP IHS module (e.g., MMS4A00-XM), not separate physical products.

The one rule this matrix proves: IHS twin-port modules go only in switches, single-port RHS modules work in both Spectrum-4 and the ConnectX-8 NIC, and QSFP112 is a separate world reserved for ConnectX-7 and BlueField-3 — so always match the module's form factor to the platform before you order.

ConnectX-7 and BlueField-3 use QSFP112 for 400G NDR — see sections 7 and 9 for full details. "—" indicates not applicable for 800G XDR modules.

13. Spectrum-4 Module Details

The Spectrum-4 supports both IHS and RHS OSFP modules, making it the only NVIDIA platform with full flexibility across the entire 800G transceiver catalog. IHS modules — including twin-port 2xDR4 and 2xFR4 — are accepted in the Spectrum-4's taller cages alongside single-port RHS modules. This dual-cage-type support is a deliberate design choice that allows Spectrum-4 to serve as the interconnect hub between ConnectX-8 NICs (which require RHS modules) and Quantum-X800 switches (which require IHS modules) within the same SuperPOD fabric.

XDR-generation links operate at 4x200G-PAM4 per lane — this is what enables the 800G per-port throughput. For breakout deployments, the Spectrum-4 supports all four breakout modes: 1x800G native, 2x400G, 4x200G, and 8x100G. Each breakout mode is controlled by the breakout cable and transceiver selection — the switch itself does not require separate configuration for the breakout ratio beyond the standard port-mode setting in Cumulus Linux or NVIDIA OS.

14. Q3400-RA Module Details

The Quantum-X800 Q3400-RA achieves its 144 effective 800G port count through twin-port IHS modules that deliver two independent 800G InfiniBand XDR links from a single physical OSFP cage. The MMS4A00-XM is the primary compatible transceiver: a twin-port 2xDR4/DR8 module in OSFP224 IHS (Closed Finned Top) form factor, providing 1.6T aggregate bandwidth per module (2×800G) over two independent MPO-12/APC single-mode fiber paths to two separate 800G endpoints.

The IHS-only cage design of the Q3400-RA is a thermal necessity. Twin-port modules dissipate significantly more heat than single-port modules — the integrated heatsink on IHS modules provides the thermal headroom required for reliable operation at 800G per port. Attempting to use RHS modules in the Q3400-RA is not a configuration error that can be worked around; it is a physical incompatibility.

15. Vitex NVIDIA-Compatible Portfolio

Vitex stocks 800G OSFP modules in both IHS and RHS form factors compatible with the full NVIDIA platform lineup. For guidance on which interconnect type pairs with each platform, see our Interconnect Selection Guide.

Use this as a quick ordering shortlist: the IHS modules (DR8, 2xDR4, 2xFR4) cover the switch tier across Spectrum-4 and both Quantum-X800 models, while the RHS modules (SR8, single DR4) pair Spectrum-4 with the ConnectX-8 NIC — pick the row that matches both your platform and your reach.

Quick-reference summary for NVIDIA 800G transceiver form factor selection across all six platforms. Keep this as the one-page reference before finalizing any transceiver order.

16. Getting Expert Support

Vitex has been a trusted fiber optics partner for over 23 years, serving data center operators, telecom carriers, and enterprise networks worldwide. With US-based engineering support and shorter lead times than major OEMs, we help teams move from design to deployment faster. Contact our engineering team for platform-specific module recommendations and compatibility verification.