Summary

What it is

NVIDIA's InfiniBand XDR platform — Q3400-RA switch plus ConnectX-8 SuperNICs — doubling per-port bandwidth to 800 Gbps with 144 ports per switch and 9× more in-network compute (14.4 Tflops via SHARPv4) than the NDR generation.

Where it fits

The fabric under the 2026 rack-scale deployments: CoreWeave's industry-first Vera Rubin NVL72 bring-up (June 1, 2026) and Azure's GB200 clusters both ride Quantum-X800. Because GPUDirect and S3-over-RDMA share this fabric, its port speed is the ceiling on what object storage can deliver into GPU HBM. The Spectrum-X800 Ethernet sibling is NVIDIA hedging the Ethernet-vs-InfiniBand split.

Misconceptions / Traps

Deployment numbers are vendor-published (CoreWeave's 10× cost-per-million-tokens claim included); independent measurements vs NDR are scarce.
XDR does not make the fabric decision for you — Dell'Oro projects Ethernet overtakes InfiniBand in AI back-ends by 2027, so this is the premium-latency fork, not the default.

Key Connections

implements RDMA (RoCE v2 / InfiniBand) — the IBTA 2.0 XDR baseline
integrates_with NVIDIA GPUDirect RDMA for S3 — the S3-to-HBM path rides this fabric
enables KV-Cache Disaggregation — sets the KV-transfer bandwidth budget
competes_with Ultra Ethernet (UEC) — the InfiniBand side of the fabric fork

Definition

What it is

NVIDIA's flagship InfiniBand XDR (eXtreme Data Rate) networking platform — the Q3400-RA switch plus ConnectX-8 SuperNICs — doubling per-port bandwidth over the previous NDR generation from 400 Gbps to 800 Gbps per port. The platform provides 144 ports of 800 Gbps connectivity and a 9× increase in in-network computing capacity (14.4 Tflops via SHARPv4) versus NDR. Its Ethernet sibling, Spectrum-X800, carries RoCEv2 with proprietary enhancements and deploys alongside InfiniBand in the same facilities at up to 1.6 Tb/s of backend bandwidth per GPU in non-blocking multi-rail topologies.

Why it exists

As GPU compute density scales, the fabric between compute nodes and storage targets became the primary bottleneck. XDR exists to keep rack-scale architectures (GB200, Vera Rubin NVL72) fed — and because S3-over-RDMA and GPUDirect Storage ride the same fabric, every doubling of port bandwidth directly raises the ceiling on what an object-storage tier can deliver to GPU HBM.

Primary use cases

AI back-end networks for rack-scale GPU architectures, KV-cache transfer between disaggregated prefill/decode pools, GPUDirect RDMA paths from S3-compatible storage into GPU memory, in-network reduction via SHARPv4 for distributed training.