NVIDIA Mellanox Spectrum & Quantum Switching Maintenance
side by side.

Spectrum SN2000/SN3000 · older mid-range Ethernet

SN2010/SN2100/SN2410/SN2700 (10/25G) · SN3000 series (25/100G)

Spectrum SN4000/SN5000 · current generations

SN4000 series (100/200G) · SN5000 series (200/400G)

Quantum InfiniBand · HPC standard

Quantum-1 (HDR 200G, QM8700) · Quantum-2 (NDR 400G, QM9700)

older Mellanox SX/SB · EOSL

SX1012/SX1024/SX1036 (Ethernet, 10/40G) · SB7700/SB7800 (InfiniBand FDR/EDR)

Generation status of Mellanox switching line

Mellanox switching maintenance has two economic constellations: current Spectrum/Quantum as TPM entry after DGX warranty expiration (synchronous AI cluster stack transition), and EOSL SX/SB as pure TPM maintenance where NVIDIA support is no longer available. For a Spectrum SN5600 (400G-capable) with 24×7×4, TechCare is typically roughly 55% below the NVIDIA support price; for the Quantum QM9700 (400G NDR) roughly 55–60%. For large AI cluster builds with 8-32 Spectrum switches plus Quantum backbone, maintenance savings become a substantial annual lever.

We service Mellanox switches with OEM switches, power supplies and optics from our own warehouse. Spectrum SN series (all four generations) and Quantum QM series (Quantum-1 and Quantum-2) are in active component pool. Older Mellanox SX/SB switches (pre-NVIDIA acquisition) we build from certified refurbishing sources — this generation is EOSL but still widespread in DACH HPC fleets. MLNX-OS on switches continues license-free — all switching functions (RoCE, EVPN, RDMA-over-Converged-Ethernet, InfiniBand SubnetManager) stay active. Code updates to newer MLNX-OS versions require an active NVIDIA contract but are usually uncritical with stable AI/HPC cluster configurations.

The full OEM-vs-TPM matrix (spare parts, SLA tiers, language, sub-outsourcing, compliance docs, reporting) is the same for every Nvidia line — see the Nvidia battle card.

BAIT/MaRisk documentation package, DORA Art. 28 contract template, ISO 27001 and GDPR-compliant media handling apply identically to every Nvidia line — details on the Nvidia battle card.

Same four steps for every Nvidia line: inventory → fixed-price quote in 48 h → contract (BAIT/MaRisk/DORA hooks) → hand-over with spare-parts provisioning. Details on the Nvidia battle card.

Which Mellanox switch models do you service?

Complete switching family: Spectrum SN series (SN2010, SN2100, SN2410, SN2700 with 10/25G; SN3000 series with 25/100G; SN4000 series with 100/200G; SN5000 series with 200/400G), Quantum InfiniBand (Quantum-1 HDR 200G with QM8700 family; Quantum-2 NDR 400G with QM9700 family) and older Mellanox switches (SX1012/SX1024/SX1036 for Ethernet 10/40G; SB7700/SB7800 for InfiniBand FDR/EDR). Including all power supplies, fan modules and SFP+/QSFP+/QSFP28/QSFP56/QSFP-DD optics modules.

What does TPM cost for Mellanox switching compared to NVIDIA support?

30 to 60 percent savings. For Spectrum SN3700 (32x100G), SN4600 (64x200G) and Quantum QM8700 (HDR 200G) with 24×7×4, TechCare is typically roughly 55–60% below the NVIDIA price; for SN5600 (64x400G) roughly 55% and for QM9700 (NDR 400G) roughly 55–60%. For an AI cluster configuration with 16 Spectrum switches plus 4 Quantum switches, TechCare is roughly 55% below NVIDIA overall. For EOSL SX/SB: up to 70 percent savings (NVIDIA often no longer offers support).

When does Spectrum Ethernet vs. Quantum InfiniBand make sense?

Spectrum (Ethernet with RoCE support) is standard for AI cluster backbones, hyperconverged data center fabrics and cloud-native workloads — RoCE brings RDMA performance on Ethernet standard networking, reducing operations complexity (Ethernet tooling, existing operations skills). Quantum (InfiniBand) is standard for classic HPC clusters — lower latency than Ethernet, integrated SubnetManager, standard in scientific computing. From AI cluster perspective: modern LLM training clusters increasingly use Spectrum (NVIDIA's own recommendation for hyperscale AI builds), classic HPC compute stays Quantum. From TPM perspective: both classes with same condition logic.

Do MLNX-OS, RoCE and InfiniBand SubnetManager continue to work without an NVIDIA contract?

Yes. MLNX-OS on Spectrum and Quantum switches continues license-free — all switching functions, RoCE (RDMA over Converged Ethernet), DCQCN (Data Center Quantized Congestion Notification), MLAG, EVPN-VXLAN on Spectrum, plus integrated SubnetManager on Quantum for InfiniBand topology management stay functionally active. NVIDIA Cumulus Linux (for Spectrum as alternative OS option) also continues license-free on hardware. MLNX-OS code updates to newer versions require an active NVIDIA contract but are usually uncritical with stable AI/HPC cluster configurations.

Can you service older SX/SB switches?

Yes. Mellanox SX series (SX1012 12-port 40G, SX1024 36-port 10G, SX1036 36-port 40G) and SB series (SB7700 36-port FDR InfiniBand, SB7800 36-port EDR InfiniBand) are still widespread in DACH HPC fleets, EOSL reached. NVIDIA often no longer offers standard support — TPM is the only remaining maintenance option. We maintain refurbishing pools for SX/SB switches and critical components (power supplies, fan modules). Pool depth varies per model — SX1036 abundantly available, SB7800 somewhat tighter but structured available. We create a risk assessment per model before contract conclusion.

Which SLA levels do you recommend for Mellanox switching?

Spectrum switches as AI cluster backbone: 24×7×4 with German-speaking onsite engineer is standard — switch outage affects AI cluster connectivity, has direct training outage impact. Quantum InfiniBand switches as HPC backbone: 24×7×4 mandatory due to InfiniBand topology criticality (SubnetManager change on switch failure). With dual-redundant spine designs (standard in modern AI clusters), single spine failure can be absorbed — 5×9 NBD per switch theoretically possible, but we recommend 24×7×4 due to cluster performance risks. For EOSL SX/SB switches in stable HPC configurations or as backup hardware: 5×9 NBD or Parts Only sufficient.