Hybrid Infrastructure

NVIDIA Networking (Mellanox).

The interconnect layer of AI — ConnectX, Spectrum-X Ethernet, Quantum InfiniBand and BlueField DPUs.

Part of NVIDIA · Also known as: Mellanox, ConnectX, Spectrum-X

Explore the portfolio Independent editorial profile

The fabric decision

Quantum InfiniBand

Spectrum-X Ethernet

From Mellanox to NVIDIA Networking

  1. Mellanox

    Mellanox Technologies builds InfiniBand and high-performance Ethernet interconnects.

  2. Acquisition

    NVIDIA acquires Mellanox; the portfolio becomes NVIDIA Networking.

  3. Today

    Interconnect is designed as part of NVIDIA's accelerated computing platform.

InfiniBand versus Spectrum-X

ConsiderationQuantum InfiniBandSpectrum-X Ethernet
HeritageHPC and AI supercomputingEthernet optimized for AI
Operations modelSpecialist InfiniBand skillsFamiliar Ethernet operations
EcosystemTightly integrated fabricBroader Ethernet ecosystem

Interconnect stack

Infrastructure offload

BlueField DPUs

Fabric

Spectrum-X Ethernet · Quantum InfiniBand

Server

ConnectX adapters and SuperNICs

Designing with compute

Interconnect architecture is co-designed with compute: topology, rail design, congestion strategy and the InfiniBand-versus-Ethernet decision per workload class. The fabric is best treated as a first-class architecture deliverable, not a cabling schedule.

Scenarios

ScenarioProblemApproachOutcome
AI fabric engineeringA GPU cluster underperforms because interconnect was an afterthought.Co-design Spectrum-X or Quantum fabric with the compute and storage, validated against workload profiles.Cluster performance that matches the silicon you paid for.
Infrastructure offloadHost CPUs burn cycles on networking, security and storage work.Deploy BlueField offload for infrastructure services.Host capacity returned to workloads; infrastructure functions hardened by isolation.
Storage-data-path performanceTraining stalls on data ingest.Engineer the end-to-end data path: NICs, fabric, storage clients.Sustained throughput that keeps accelerators fed.

Evaluation considerations

Organizations evaluating NVIDIA Networking typically treat AI interconnect as part of whole-system architecture: fabric topology, workload-validated design, and DPU offload strategy. Performance validation before production commitment is the key adoption safeguard given how much cluster performance depends on the interconnect layer.

Questions buyers ask

  1. Q1Which fabric suits our operations team?
  2. Q2Are we training at scale or serving inference?
  3. Q3How will performance be validated before commitment?

FAQ

InfiniBand or Ethernet for AI?

Quantum InfiniBand for maximum training performance; Spectrum-X where Ethernet operations models and ecosystem matter. Both are NVIDIA-engineered for AI; workload profile and team skills decide.

Is Mellanox still a brand?

No — the portfolio is fully NVIDIA Networking. Existing Mellanox-era equipment remains widely deployed and supported within the NVIDIA structure.

What is a DPU actually for?

Moving infrastructure work — networking, security, storage services — off host CPUs onto dedicated silicon. The payoff is workload capacity and stronger isolation.

Is this needed for inference-only estates?

At scale, yes — inference serving has its own demanding traffic patterns. For modest inference, standard networking suffices; sizing should follow the real workload profile.

Official further reading

Independent editorial profile. Vendor facts reviewed against official sources, September 24, 2026.

Related technologies

Discuss your technology priorities.

Planning changes across data, AI, cloud or infrastructure? Tell us about your priorities.

Contact us