NVIDIA Mellanox ConnectX & BlueField Adapter Maintenance
side by side.

ConnectX-4 / ConnectX-5 · older generations

ConnectX-4 (25/40/100G EN+IB) · ConnectX-5 (25/100G EN+IB)

ConnectX-6 Dx · current mid-range

ConnectX-6 Dx (25/100/200G) · ConnectX-6 VPI

ConnectX-7 · current high-end

ConnectX-7 (200/400G EN+IB)

BlueField DPU · data processing units

BlueField-2 (25/100G with ARM cores) · BlueField-3 (200/400G with extended ARM compute)

Generation status of Mellanox adapters

Mellanox adapter maintenance has two economic sweet spots: volume maintenance for ConnectX NICs and high-value maintenance for BlueField DPUs. ConnectX NICs are relatively cheap per unit but deployed in high quantities in AI/HPC cluster builds — 100-1000+ adapters are not uncommon. NVIDIA standard support or distributor maintenance is billed per adapter per year. For 200+ adapters, this adds up to a significant budget item — TPM reduces this to 30-50 percent. BlueField DPU is calculated differently: high hardware value per unit, NVIDIA support proportionally more expensive, TPM pays off from first unit. For mixed fleets (NICs + DPUs) we offer multi-class contracts with different pricing logics.

We service Mellanox adapters with OEM components from our own warehouse. ConnectX-4 and ConnectX-5 (older generations, high install density in DACH) are abundantly available. ConnectX-6 Dx and ConnectX-7 (current generations) we continuously build the component pool. BlueField-2 and BlueField-3 included as high-value components in our pool. MLNX_OFED drivers and software stack (RDMA, GPUDirect, NVMe-oF, OVS-Offload, ASAP² for BlueField) continue license-free — all driver functions stay active under TPM. Software subscriptions like NVIDIA AI Enterprise or DOCA SDK are independent of hardware maintenance.

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 adapter models do you service?

Complete adapter family: ConnectX-4 (25/40/100G Ethernet and InfiniBand), ConnectX-5 (25/100G EN+IB), ConnectX-6 Dx (25/100/200G), ConnectX-6 VPI (Virtual Protocol Interconnect with dual EN/IB), ConnectX-7 (200/400G EN+IB) plus BlueField-2 DPU (25/100G with ARM cores) and BlueField-3 DPU (200/400G with extended ARM compute platform). Including all PCIe form factors (standard, OCP 2.0, OCP 3.0) and optics modules for respective speeds.

From how many adapters does TPM pay off for Mellanox?

Rule of thumb: from 50 ConnectX NICs (same or mixed generation) volume TPM becomes economically attractive — contract administration, SLA reporting and onsite engineer readiness justify minimum contract value. For 100+ NICs annual savings are significant. For BlueField DPUs TPM pays off from first unit because hardware value per DPU is high and NVIDIA support proportionally more expensive. Recommendation: consolidate mixed fleets (e.g. 200 ConnectX-5 + 20 BlueField-3) in one multi-class contract.

What does TPM cost for Mellanox adapters concretely?

ConnectX NIC maintenance: depending on generation and SLA, TechCare is roughly 50–55% below NVIDIA support or distributor maintenance — including for 200 ConnectX-5 adapters. BlueField DPU: roughly 55–60% below NVIDIA premium support (depending on generation and SLA), e.g. for 20 BlueField-3 DPUs. Mixed fleet 200 NICs + 20 DPUs: roughly 55% below the NVIDIA price overall.

Do MLNX_OFED, RDMA, GPUDirect and DOCA continue to work without an NVIDIA contract?

Yes. MLNX_OFED drivers and software stack continue license-free on the adapters — RDMA (RoCE/InfiniBand), GPUDirect (for AI clusters), GPUDirect Storage, NVMe-over-Fabric, ASAP² (for BlueField), Open vSwitch offload and all standard driver functions stay functionally active. MLNX_OFED updates to newer versions are freely available from NVIDIA — TPM customers can download them without restriction. DOCA SDK (for BlueField programming) and NVIDIA AI Enterprise are software subscriptions, independent of hardware maintenance — customers actively using them keep them with NVIDIA. Firmware updates on adapters require an active NVIDIA contract but are usually uncritical with stable configurations.

Which SLA levels do you recommend for Mellanox adapters?

Strongly depends on cluster architecture: ConnectX NICs in productive AI/HPC cluster nodes with redundant multi-path setups (standard in modern AI build): 5×9 NBD economically appropriate — single NIC failure absorbed by multi-path. BlueField DPUs as central storage offload or security inspection components: 24×7×4 mandatory due to service impact on failure. For single NIC setups (rare in modern AI builds) or critical database storage targets: 24×7×4 mandatory. For fleet migration during cluster refresh: Parts Only sufficient for out-phase adapters.

Can you service older ConnectX-3 or pre-Mellanox-acquisition fleets?

ConnectX-3 (released 2012) and ConnectX-3 Pro (2014) are completely EOSL and practically no longer productive in the market. If fleets are present, we can check coverage on request — typically refresh is the only sensible option here. ConnectX-2 and older are no longer part of our pool. Mellanox acquisition by NVIDIA was 2020 — all adapter models from ConnectX-4 are today under NVIDIA branding and serviced by us regularly, regardless of original manufacturer branding (Mellanox before 2020, NVIDIA since 2020).