OSFP DAC Cables Guide

→ Browse the DAC Cables collection

Companion reading: OSFP DAC Cables: What They Are, When to Use One, and When They'll Quietly Ruin Your Week

OSFP DAC Cables — 400G NDR, 800G & Breakout Direct Attach Copper

This guide covers Octal Small Form-factor Pluggable (OSFP) direct attach copper assemblies from 200G through 1.6T, in passive, active copper, and breakout configurations, for NDR/XDR InfiniBand and 400G/800G Ethernet fabrics. It is written for engineers specifying cable BOMs and assumes you already know what a DAC cable is. For the plain-English introduction, read the companion blog article linked above.

1. OSFP at a Glance

2. Lane Architecture by Generation

The connector shape tells you very little about the link rate. Two assemblies can both be "OSFP" and be entirely different products. Always match the part number to the port.

The twin-port distinction matters when specifying NVIDIA fabrics. A Quantum-2 switch port presents 800G as two independent 400G ports behind a single physical OSFP cage, which is why so much NDR cabling is breakout rather than straight-through.

3. Cable Class Selection: Passive vs ACC vs AEC vs AOC

Reach figures are indicative planning values, not guarantees. Validate against the host and switch vendor's published supported-cable list for the specific part number before committing a BOM.

4. Why Reach Shrinks as Lane Rates Rise

Passive copper offers no regeneration anywhere along the link. The receiver must recover the signal from whatever survives the channel, and the available budget is consumed by conductor insertion loss (which rises with frequency and length), dielectric loss, connector and PCB losses at both ends, and crosstalk between adjacent differential pairs.

Moving from 50G-PAM4 to 100G-PAM4 roughly doubles the Nyquist frequency, which materially increases loss per metre while the available budget stays broadly similar. Heavier gauge copper buys back some length at the cost of stiffness and bulk — which is why a 3 m 400G DAC is noticeably thicker and less flexible than a 1 m one.

FEC is assumed, not optional, at 100G-per-lane rates. A link that is "up" but reporting a rising corrected-codeword rate is running with reduced margin and should be treated as marginal.

5. Mechanical and Thermal Variants: Flat Top vs Finned Top

Flat top — smooth module housing, used where the receptacle cage provides the cooling. Standard on host adapter ports including ConnectX-7.

Finned top — integrated heatsink fins on the module itself, used in switch cages designed around module-supplied cooling. Standard on NVIDIA Quantum-2 switch ports.

A cable's end configuration must be specified as a pair:

• Flat / flat — host-to-host (for example, a direct DGX-to-DGX link)
• Flat / finned — host-to-switch
• Finned / finned — switch-to-switch

This is not electrically significant but it is mechanically mandatory: the wrong end type will not seat in the cage. Also observe the published minimum bend radius for the specific assembly. At 26 AWG the cable is stiff, and violating bend radius near the connector is a common cause of intermittent links that pass a bench test and fail once dressed into a rack.

6. Breakout Topologies

Breakout is the default for connecting Quantum-2 switches to ConnectX-7 hosts, because each switch cage fronts two 400G ports. Budget port counts on the delivered side, not the physical cage count, or you will under-order switches.

7. NVIDIA / Mellanox Part Number Decoding

So MCP4Y10-N001 reads as: passive copper, single-port OSFP family, NDR InfiniBand, 1 metre. Treat this as a reading aid rather than a specification — always confirm the full part against NVIDIA's official LinkX parts list.

8. Worked Example: MCP4Y10-N001 (400G NDR, 1 m)

The clearest real-world OSFP DAC deployment is a switch-free two-node NDR InfiniBand link between NVIDIA DGX H100 or H200 systems.

Topology. Each DGX H100/H200 exposes its ConnectX-7 cluster fabric through OSFP ports on the rear of the chassis. One cable end into node A, the other into node B, and you have a direct 400 Gb/s NDR link with no switch, no transceivers, and no adapters. NCCL collectives run across it directly.

Constraints to respect. This topology terminates at two nodes — three or more requires a Quantum-2 InfiniBand switch. The ports run InfiniBand, not Ethernet. Use OSFP, never QSFP-DD: 400GbE parts such as the MCP1660-W series are Spectrum Ethernet cables and will not carry NDR on these ports. Choose 1 m for adjacent placement, or MCP4Y10-N002 at 2 m when the nodes sit further apart in the rack.

→ Browse OSFP and NDR DAC cables in the DAC Cables collection

9. Platform Notes

Support is ultimately governed by adapter and switch firmware. Firmware revision, not the cable, is frequently the reason a physically correct part fails to link.

10. Installation and Validation

Seat the cable fully until the latch engages audibly, confirm the end type matches the cage, and route within the published bend radius before dressing the cable. Then validate rather than assume:

• ibstat — confirm port state, physical state, and negotiated rate on InfiniBand links
• mlxlink -d <device> -m — read module and cable identification from the adapter
• mlxlink -d <device> -e — review effective BER and error counters
• ibdiagnet — fabric-wide sweep for link errors and topology mismatches
• ethtool -m <interface> — module identification on Ethernet ports

A healthy passive DAC link shows the expected rate, a stable physical state across a thermal soak, and a corrected-codeword rate that is flat over time rather than climbing.

11. Troubleshooting Matrix

12. Ordering Guidance

Measure the routed path, not the straight-line distance between ports, and add slack for service loops. Specify end types explicitly. Confirm the fabric — InfiniBand and Ethernet parts are not interchangeable even at identical rates. Where the routed length lands within 20% of the passive ceiling for your lane rate, specify active copper instead; the cost delta is far smaller than the cost of diagnosing an intermittent fabric.

Frequently Asked Questions

Is OSFP the same as QSFP-DD? No. Both carry eight lanes and both support 400G and 800G, but the connectors are mechanically different and not interchangeable. QSFP-DD is backward compatible with QSFP; OSFP is physically larger with greater thermal headroom. An OSFP DAC will not fit a QSFP-DD cage.

How far can a 400G passive OSFP DAC run? Plan for roughly 1 to 2 metres, occasionally 3 metres depending on the assembly and platform. Beyond that, move to ACC, AEC, or optical.

Does a passive OSFP DAC consume power? No meaningful amount. There is no optical engine and no retimer, so the only draw is the host port itself.

Can I use a 400GbE QSFP-DD DAC on a DGX H100? No. The DGX H100/H200 cluster fabric uses OSFP ports running NDR InfiniBand. QSFP-DD 400GbE parts are for Spectrum Ethernet switches.

Why is my 400G link up but reporting FEC corrections? The link is running with reduced margin. Check cable length against the platform ceiling, verify bend radius, and inspect for thermal drift before blaming the driver.

Can I connect three or more DGX H100 nodes with DACs alone? No. Direct DAC links work for two nodes. Three or more requires a Quantum-2 InfiniBand switch, at which point you move to breakout cabling.

How do I confirm a cable is validated for my switch? Check the vendor's official supported-parts list and buy from a supplier that programs and tests the EEPROM. Ours are pre-coded so the port recognises them at boot with no manual intervention.

The Resilient Tec Guarantee

• ✅ Pre-programmed EEPROM — recognised immediately by your switch, no errors on boot
• ✅ Compatibility-tested in our lab before every shipment
• ✅ Lifetime warranty on all DAC cables
• ✅ Talk to a real engineer, not a chatbot

Not sure which OSFP part your topology needs? Contact us and we will confirm compatibility for free before you buy.