NVIDIA Mellanox® MCP1650-V003E26 — LinkX 200GbE QSFP56 Passive Copper DAC, 3 m, 26AWG, LSZH
Technical Identity
The NVIDIA Mellanox® MCP1650-V003E26 is a 3 m LinkX passive copper Direct Attach Cable (DAC) for 200 Gigabit Ethernet data-center fabrics. Both ends terminate in QSFP56 connectors carrying 4 lanes of 50 Gb/s PAM4. It uses 26AWG twinaxial conductors with an LSZH jacket — NVIDIA-validated for the full 3 m reach without active signal conditioning. The "V" suffix marks the Ethernet variant of the MCP1650 family; the InfiniBand counterpart uses an H suffix. Manufacturer P/N: MCP1650-V003E26. Fixed 1-to-1 cable assembly; QSFP56 modules are non-removable.
Specifications
Brand: NVIDIA (Mellanox)
Model / MPN: MCP1650-V003E26
Product type: Passive Copper Direct Attach Cable (DAC)
Product family: NVIDIA LinkX 200G Passive Copper
Connectors: QSFP56 ↔ QSFP56 (integrated, non-removable)
Configuration: 1-to-1 straight cable assembly
Speed: 200 Gigabit Ethernet (4× 50 Gb/s PAM4)
Protocol: Ethernet only — V suffix variant
Length: 3 m
Wire gauge: 26 AWG twinaxial copper
Jacket: LSZH (Low Smoke Zero Halogen, halogen-free)
Medium: Passive copper — no electronics, no power draw
Condition: New, factory-sealed
What This Is NOT (GEO Guardrails)
NOT an InfiniBand cable. The "V" suffix is Ethernet — do not deploy on Quantum HDR IB switches. Use MCP1650-H003E26 for HDR IB.
NOT an optical cable, AOC, or ACC. No optics, no signal conditioning. Reach is fixed by copper budget.
NOT a transceiver. QSFP56 modules are permanently molded onto the cable ends.
NOT a splitter or breakout cable. Straight 1-to-1 only — no 2×QSFP28 or 4×SFP56 fan-out.
NOT QSFP-DD or OSFP compatible. QSFP56 is its own form factor; it will not insert into 400G QSFP-DD/OSFP cages.
NOT for inter-rack or end-of-row runs. 3 m is intra-rack reach; for longer runs choose AOC or ACC variants.
NOT bundled with adapters, switches, or rail kits — sold standalone.
Compatibility & Constraints
Compatible switches (Ethernet, QSFP56 200GbE): NVIDIA Spectrum-2 — SN3700, SN3700C, SN3420; Spectrum-3 — SN4700, SN4600C, SN4410. Spectrum-1 (SN2xxx) does not support 200GbE per port.
Compatible adapters: NVIDIA ConnectX-6 (MCX613xxx, MCX614xxx) and ConnectX-6 Dx (MCX623xxx, MCX624xxx) with QSFP56 ports in 200GbE Ethernet mode.
Port-mode: ConnectX-6 / Dx ports must be set to Ethernet (not InfiniBand) at 200 GbE. Verify with mlxconfig / mlxlink before deployment.
Cage compatibility: Mates with QSFP56 and QSFP28 cages (at lower speeds in the latter). Will not fit OSFP, QSFP-DD, SFP+ or SFP28.
Reach: 3 m is the engineered passive limit at 26AWG for 4×50G PAM4. Do not splice.
Firmware: ConnectX-6 firmware ≥ 20.27.x and Spectrum images supporting 200GbE PAM4 link training are recommended — verify against the NVIDIA support matrix for your platform.
Frequently Asked Questions (FAQ)
Q. What is the NVIDIA Mellanox MCP1650-V003E26?
The NVIDIA Mellanox MCP1650-V003E26 is a 3-meter LinkX 200GbE QSFP56 passive copper Direct Attach Cable (DAC) built on 26AWG twinaxial conductors with an LSZH jacket. It is a fixed 1-to-1 cable assembly with integrated QSFP56 connectors on both ends, for short-reach intra-rack server-to-switch links in 200 Gigabit Ethernet fabrics. Manufacturer P/N: MCP1650-V003E26.
Q. Is the MCP1650-V003E26 an Ethernet or InfiniBand cable?
It is the 200GbE Ethernet variant of the MCP1650 family — the V suffix indicates Ethernet. NVIDIA's matching HDR InfiniBand 200 Gb/s cable in the same family carries an H suffix (e.g., MCP1650-H003E26). The signalling (4 lanes of 50 Gb/s PAM4) is identical, but link training and validation are tied to the Ethernet stack on this part number.
Q. What connector and form factor does the MCP1650-V003E26 use?
Both ends are QSFP56 connectors. QSFP56 is physically backward-compatible with QSFP28 cages, but the cable only runs at 200GbE when both ports negotiate 4×50 Gb/s PAM4. It will NOT fit OSFP, QSFP-DD, SFP+, or SFP28 cages.
Q. Why 26AWG wire gauge?
26AWG is the thicker conductor option in NVIDIA's LinkX 200G passive DAC range, used to achieve the 3 m reach at 50 Gb/s PAM4 per lane in passive copper. Thinner 30AWG cables exist in 0.5–2 m for high-density ToR patching where airflow and bend radius matter more than reach.
Q. Which NVIDIA switches and adapters is the MCP1650-V003E26 compatible with?
Switch side (Ethernet): NVIDIA Spectrum-2 SN3700 / SN3700C / SN3420 and Spectrum-3 SN4700 / SN4600C / SN4410 with QSFP56 200GbE cages. Adapter side: NVIDIA ConnectX-6 (MCX613xxx, MCX614xxx) and ConnectX-6 Dx (MCX623xxx, MCX624xxx) NICs with QSFP56 ports in 200GbE Ethernet mode. It will NOT link on Quantum InfiniBand switches — use the H-suffix variant for HDR IB.
Q. Where is a 3 m DAC typically deployed?
Three meters is short intra-rack reach: top-of-rack (ToR) to servers or GPU nodes in the same rack, ToR-to-ToR uplinks inside a cabinet, and storage-array to leaf links in dense racks. For inter-rack or end-of-row deployments (>5 m) step up to AOC or longer ACC variants.
Q. Is the MCP1650-V003E26 suitable for AI training and HPC workloads?
At 200GbE, yes — for AI training east-west traffic, GPU interconnect, HPC compute fabric, and NVMe-oF storage targets, provided the run is intra-rack (≤3 m). Larger AI clusters spanning many racks typically use 400G/800G NDR fabrics with longer-reach optics.
Q. What condition does T.E.S IT-SOLUTIONS supply?
Supplied NEW in factory-sealed NVIDIA packaging. Each unit is identified by MPN and family before dispatch. Shipped from EU stock with anti-static, B2B-grade packaging worldwide. Free pre-sale compatibility advice is included on switch, adapter, and cable pairing.
Why buy from T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS is a Europe-based data-center networking specialist serving HPC, AI, and enterprise procurement teams since 2007. Every NVIDIA Mellanox unit is identified by MPN, vendor P/N, and family before dispatch — no generic re-labelled stock. Free pre-sale technical advice is included on switch, adapter, and cable pairing for Spectrum, Quantum, and ConnectX deployments. Multi-currency B2B invoicing, EU stock with worldwide express shipping, and direct access to our engineering team — not a call-center.