Short answer: An OSFP DAC is a fixed-length copper cable with OSFP connectors moulded onto both ends. It carries 400G or 800G over roughly one to three metres, draws essentially no power beyond the port, and costs a fraction of optics. Use it inside a rack or between adjacent racks. Go further than that and you want active copper or an optical cable instead.
What is an OSFP DAC cable?
OSFP stands for Octal Small Form-factor Pluggable — "octal" because the connector carries eight electrical lanes. A DAC (Direct Attach Copper) cable takes those lanes, runs them down shielded twinax copper, and permanently attaches an OSFP connector housing to each end. No transceiver to buy. No fibre to clean. No optical engine to fail. You unbox it, you plug it in, the link comes up.
That is the whole pitch, and it is a good one: copper is cheap, copper is fast, copper needs no power. The catch is that copper is also physically honest about distance in a way fibre simply is not.
How does copper carry 400G?
By splitting the job. A 400G NDR InfiniBand OSFP link runs four lanes at 100 Gb/s each using PAM4 signalling, which encodes two bits per symbol instead of one. An 800G twin-port OSFP uses all eight lanes at 100G. The newest generation moves to 200G-class lanes to reach 1.6T.
The practical consequence: the connector shape tells you very little about the speed. Two cables can both be "OSFP" and be completely different products. Always match the part number to the port, not the plastic.
How far will an OSFP DAC actually go?
This is where people get burned. Higher lane rates mean less loss budget, which means shorter cables.
| Link rate | Typical passive OSFP DAC reach | Beyond that, use |
|---|---|---|
| 200G (2 × 100G) | ~3 m | ACC / AOC |
| 400G NDR (4 × 100G) | ~1–2 m, sometimes 3 m | ACC / AEC / AOC |
| 800G (8 × 100G) | ~1–2 m | ACC / AEC / AOC |
| 1.6T (200G lanes) | ~1 m or less | Active copper / optics |
These are planning numbers, not guarantees. The authoritative answer is always your switch or NIC vendor's validated parts list for the exact part number. The loss budget behind these figures is covered in the OSFP DAC Cables Guide.
Flat top or finned top? The detail nobody warns you about
OSFP modules come in two physical flavours, and mixing them up means the cable does not seat.
Finned top has an integrated heatsink on the module and generally goes into switch cages that rely on the module's own fins for airflow — think NVIDIA Quantum-2 switch ports. Flat top is smooth and goes into host adapter ports where the cage supplies the heatsink — think ConnectX-7 on a DGX.
So a DGX-to-DGX cable is flat top on both ends. A DGX-to-switch cable is usually flat on one end and finned on the other. Order the wrong combination and you own a very expensive bookmark.
Five scenarios where an OSFP DAC is the right call
1. Two DGX nodes, no switch
This is the cleanest use case there is. Two NVIDIA DGX H100 or H200 systems, each with ConnectX-7 OSFP ports on the back, linked directly at 400G NDR with a single passive cable. No Quantum-2 switch, no transceivers, no fabric manager gymnastics. The MCP4Y10-N001 — single-port OSFP to OSFP, 400G NDR, 1 m — is exactly this cable. One metre suits nodes in the same or an adjacent rack position; step up to the 2 m MCP4Y10-N002 if they sit further apart. Note that this trick stops at two nodes: three or more and you are buying a switch.
2. Top-of-rack to host, GPU pod style
Server in the rack, leaf switch at the top of the same rack, run under two metres. Passive copper is cheaper, cooler and lower-latency than anything else you could put there. Across an eighty-node pod, the power saved on optics alone is real money.
3. Switch-to-switch inside the same rack
Stacking two leaf switches or cross-linking an MLAG pair. Short, fixed, never moves. Perfect DAC territory.
4. Breakout fan-out
An 800G twin-port OSFP switch port split into two 400G or four 200G host links. One cable, one port, multiple servers — a genuinely large saving in switch port cost and in cable management. We stock these, including 800G-to-4x200G and 800G-to-2x400G configurations.
5. Lab, staging and burn-in racks
Environments where you rewire constantly and do not want to risk optics. Copper shrugs off abuse that would take a fibre patch out of service.
→ Browse the DAC Cables collection
When an OSFP DAC is the wrong answer
If the run crosses a row, goes up a floor, or has to be routed through overhead trays with tight bends, stop. Passive copper at 400G is thick, stiff and unforgiving of bend radius, and you are past its loss budget anyway. Reach for active copper for the awkward middle distances, and optics for anything genuinely long. Also skip DAC if you need to re-length cables often — every DAC is a fixed length, and "close enough" cable management at 400G is how you end up chasing intermittent CRC errors at 2 a.m.
The failure mode to recognise
An over-length or marginal passive DAC rarely fails cleanly. It negotiates down, or links and flaps under thermal load, or holds a link while quietly logging FEC corrections and retransmits. If a 400G link is "up but weird", measure the cable length before you blame the driver.
FAQ
Is OSFP the same as QSFP-DD? No. Both carry eight lanes and both do 400G and 800G, but the connectors are mechanically different and not interchangeable. QSFP-DD is backward compatible with QSFP; OSFP is physically a bit larger with better thermal headroom. An OSFP DAC will not go into a QSFP-DD cage.
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 such as the MCP1660-W series are Ethernet cables for Spectrum switches and will not do the job.
Does a passive OSFP DAC need power? No meaningful amount. There is no optical engine and no retimer, so the only draw is the host port itself. That is the main reason large clusters use copper wherever the geometry allows.
Passive, ACC or AEC — what is the difference? Passive is bare copper. ACC (active copper cable) adds signal conditioning to stretch the reach a little. AEC (active electrical cable) adds full retimers for the longest copper runs. Reach goes up, and power and cost go up with it.
Is a DAC lower latency than an AOC? Marginally, yes — there is no electrical-to-optical conversion. The difference is small in absolute terms but measurable, which is why HPC and trading environments prefer copper for short hops.
Can I mix DAC and optics in one fabric? Absolutely, and most real deployments do. Copper in-rack, optics between rows.
How do I know 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.
Go deeper
The engineering detail — lane maps, insertion-loss budgets, part-number decoding, breakout topologies and validation commands — lives in the OSFP DAC Cables Guide. If you would rather just have someone confirm the right part for your topology, talk to us; we will check compatibility before you spend anything.