NVIDIA 920-9B36F-00RX-8S0 QUANTUM-3 Based XDR Infiniband Switch Q3400-RA 4U 144 XDR Ports Over

NVIDIA 920-9B36F-00RX-8S0 QUANTUM-3 XDR InfiniBand Switch

Overview

The NVIDIA 920-9B36F-00RX-8S0 is a 4U QUANTUM-3-based XDR InfiniBand switch delivering 144 ports of high-speed fabric connectivity for enterprise data center environments. This model is purpose-built for organizations requiring dense, low-latency interconnect infrastructure to support GPU clusters, high-performance computing (HPC), and large-scale machine learning workloads where fabric performance directly impacts application throughput.

At 168.65 lb and measuring 30.50 × 46.25 × 15.50 inches, the Q3400-RA fits standard 4U rack enclosures — a critical constraint for data center retrofit scenarios where space is finite and every rack unit counts. The 144 XDR ports provide the port density needed to minimize multi-hop traffic patterns and reduce fabric latency in large-scale deployments.

Key Features

• 144 XDR ports: Each port supports XDR (Extreme Data Rate) signaling, delivering massive per-port bandwidth. In a 144-port configuration, you gain fabric scalability without sacrificing latency — critical when application performance hinges on sub-microsecond communication between compute nodes.
• QUANTUM-3 architecture: The underlying QUANTUM-3 silicon is engineered for deep buffering and intelligent congestion management. This means you can run high-utilization workloads (GPU training, large distributed databases, financial simulations) without packet loss or the kind of tail-latency spikes that break SLA guarantees on batch jobs.
• 4U form factor: Compact enough to fit in standard data center racks without custom cabling runs or external enclosures. The 30.50-inch length fits within typical rack depth limits, simplifying cable management and reducing integration labor on large fabric deployments.
• High-density interconnect: 144 ports in a single chassis eliminate the need for cascading switch tiers in smaller to mid-scale clusters. Fewer hops mean lower latency and simpler operational overhead — you're not debugging multi-layer fabric topology issues when everything is one switch away.
• Enterprise-class reliability: NVIDIA QUANTUM switches are built for continuous operation in production data centers. The switch is designed to handle full-line-rate traffic across all 144 ports without thermal throttling or performance degradation during peak compute utilization.
• Integration with NVIDIA AI/HPC ecosystems: The 920-9B36F-00RX-8S0 (often searched as 920 9B36F 00RX 8S0) is a native fit for GPU-dense environments running CUDA applications, NVIDIA DGX systems, or multi-node AI training clusters. InfiniBand fabric is the standard interconnect for NVIDIA's high-performance computing reference architectures, ensuring vendor-validated compatibility with your software stack and deployment guides.

Integration & Compatibility

The NVIDIA 920-9B36F-00RX-8S0 integrates with InfiniBand-based data center fabrics serving HPC, AI training, and real-time analytics clusters. Compatibility depends on your fabric architecture: it works as a core switch in a fat-tree or leaf-spine topology, or as a dedicated fabric switch in GPU-accelerated environments where every node must communicate with every other node at wire-speed.

Organizations running Kubernetes clusters on GPU nodes, SLURM-managed HPC workloads, or distributed machine learning frameworks (PyTorch Distributed, TensorFlow on multi-GPU setups) benefit from the low-latency, lossless characteristics of InfiniBand. If your deployment already relies on NVIDIA GPUs and requires sub-millisecond inter-node latency, the 920-9B36F-00RX-8S0 is the fabric layer you spec to avoid becoming a bandwidth bottleneck.

What's in the Box

No package contents documented in evidence. Contact the manufacturer or your distribution channel for confirmation of included cabling, documentation, and accessories.

Frequently Asked Questions

Q: What is the bandwidth per XDR port on the 920-9B36F-00RX-8S0?

A: XDR (Extreme Data Rate) InfiniBand ports deliver significantly higher bandwidth than previous-generation DDR or HDR ports. The QUANTUM-3 architecture supports the latest InfiniBand speed grades; consult the official NVIDIA datasheet for exact per-port specifications.

Q: Can the 920-9B36F-00RX-8S0 be used in a heterogeneous fabric with DDR or HDR switches?

A: InfiniBand fabrics require careful speed negotiation across mixed-generation switches. XDR ports may operate at reduced speeds when connected to older switch generations. For optimal performance and to avoid fabric bottlenecks, uniform switch generation is recommended in production deployments.

Q: What is the maximum distance for XDR InfiniBand cables from the 920-9B36F-00RX-8S0?

A: InfiniBand cable runs are limited by signaling integrity. Standard copper cables support runs up to 7 meters at XDR speeds; fiber optic cables extend this to 100+ meters. Verify cable specifications with your supplier to match your rack layout and ToR (top-of-rack) placement strategy.

Q: Does the 920-9B36F-00RX-8S0 require specialized fabric management software?

A: NVIDIA provides OpenFabrics Enterprise Distribution (OFED) stack and OpenSM (OpenFabrics Subnet Manager) for InfiniBand fabric provisioning and diagnostics. Most enterprises also deploy NVIDIA's Unified Fabric Manager or third-party fabric monitoring tools to track link health, congestion, and performance metrics across the switch.

Q: What power and cooling requirements does the 168.65 lb Q3400-RA have?

A: At full line-rate utilization across 144 ports, the switch consumes significant power and generates heat. Confirm power supply capacity (typically redundant PSUs in data center configurations) and ensure your rack cooling can handle full-duplex traffic across all ports. Power and thermal specifications are in the official NVIDIA product documentation.

Q: Is the NVIDIA 920-9B36F-00RX-8S0 NDAA Section 889 compliant?

A: NDAA compliance for this specific SKU is not confirmed in available documentation. Verify directly with NVIDIA or your procurement team if compliance is a contractual requirement for your deployment.

Marty Allison

The NVIDIA 920-9B36F-00RX-8S0 is a fabric-layer decision, not a typical network appliance. If you're building a multi-node GPU cluster or HPC environment where sustained inter-node bandwidth and latency matter more than WAN throughput, this switch is where your infrastructure either succeeds or bottlenecks. The 144 XDR ports and QUANTUM-3 silicon are engineered to keep all your compute nodes talking at wire-speed without congestion collapse.

Technical Highlights:

• 144 XDR ports in 4U: Port density of 36 ports per rack unit translates directly to fewer fabric hops in GPU training clusters. At the scale where you're running 32-node PyTorch Distributed training jobs, every additional hop adds microseconds of latency per message round-trip — this switch architecture eliminates that tax.
• QUANTUM-3 buffering and congestion control: Deep packet buffers prevent the kind of incast collisions that happen when all compute nodes try to write results back to a shared NVMe storage backend simultaneously. The switch absorbs temporary traffic bursts without dropping frames, keeping your MPI (Message Passing Interface) library from timing out on collective operations.
• Native NVIDIA ecosystem integration: The 920-9B36F-00RX-8S0 is the validated interconnect for NVIDIA DGX SuperPOD, NVIDIA HGX clusters, and any CUDA Collective Communications Library (NCCL) deployment. Your NVIDIA field engineer will hand you this SKU as part of the reference architecture — it's not an afterthought.

Deployment Considerations:

• InfiniBand is lossless, not Ethernet. If you're migrating from a Clos-topology Ethernet fabric, your operational model changes — no more TCP retransmits masking congestion. You get predictable performance, but you also get visibility into fabric saturation that forces you to right-size your cluster workloads upfront.
• Cable management at 144 ports is a physical reality. Plan for patch panel space, cable runs, and replacement procedures. An XDR copper cable costs more than a Category 6A Ethernet cable, and logistics matter at scale.

This switch is built for data centers running NVIDIA GPU workloads where the marginal cost of fabric latency exceeds the cost of the switch itself. If your cluster is smaller than 32 nodes or your workload is primarily storage-bound rather than compute-bound, you may not see ROI on XDR-generation fabric — but if you're at the scale where inter-node latency is your SLA constraint, the 920-9B36F-00RX-8S0 is the decision that unlocks your GPU cluster's full potential.