Five types of cable serve AI cluster connections. The choice is determined primarily by distance and secondarily by whether you need field-replaceability of the optical components.
Cable selection — pick your scenario
Select run distance and port speed to see the right cable type
Passive DAC -- Direct Attach Copper
A passive copper cable with OSFP or QSFP-DD connectors at both ends. No electronics inside the connectors. The electrical signal runs directly through copper wire from port to port.
Maximum distance: 7m for 400G, shorter for 800G. Power consumption: below 0.15W (essentially parasitic). Cost: the cheapest cable option by a wide margin. Latency: lowest of all options -- no optical conversion, no active components.
Use for: server-to-ToR switch connections (1-3m typical in a standard rack). This is the standard cable type for all eight DGX NIC to leaf switch connections within the same rack in a BasePOD.
Failure mode: internal copper strand break, usually from excessive bending. When a passive DAC fails, the switch port goes Err-Disabled. The NIC, however, continues reporting the port as Active -- it cannot detect the failure at the electrical level. This two-views-disagree failure mode is exactly what Lab 0 simulates. The diagnostic chain is: NCCL timeout -> ibstat shows rail down -> show interface counters on switch shows Err-Disabled -> ethtool -S on the DGX NIC shows the NIC thinks the link is up -> suspect cable.
Active Electrical Cable / AEC -- and NVIDIA's ACC, which is not the same thing
A copper cable with signal conditioning electronics in the connector heads. The electronics amplify and reshape the signal, allowing slightly longer copper runs than passive DAC at high speeds.
Maximum distance: 7m at 800G. Power consumption: 1-2W per cable. Cost: higher than passive DAC, lower than optical. Use for: intra-rack 800G connections where passive DAC cannot maintain signal integrity (at 800G, passive DAC is typically limited to 3-5m).
The distinction worth holding onto: "active copper" covers two different implementations, and NVIDIA's LinkX line uses the cheaper one.
A true AEC carries a retimer or DSP in the head -- it recovers the clock, regenerates the signal, and re-transmits it. That costs power and adds latency.
NVIDIA's ACC (Active Copper Cable) uses linear pre-emphasis ICs, not DSPs. It shapes the signal without retiming it, which is why an ACC draws about 1.5 W where an AEC draws more, and why it adds no meaningful latency. The trade is reach: ACC tops out around 5 m.
If a design document says AEC and the part number is an NVIDIA ACC, the power and latency budgets in that document are wrong in your favour -- but they are still wrong.
The lengths that actually exist
Reach is not continuous. You order from a SKU list, and the list is shorter than most rack plans assume:
| Link | Passive DAC | ACC |
|---|---|---|
| Switch to switch, 2x400G | 0.5, 0.75, 1, 1.5, 2 m | 3, 4, 5 m |
| Switch to 400G OSFP HCA (1:2 splitter) | 1, 1.5, 2, 2.5, 3 m | 4, 5 m |
| Switch to 400G QSFP112 (1:2 splitter) | 0.5 to 3 m | 4, 5 m |
| Switch to 200G (1:4 splitter) | 0.5 to 3 m | 4, 5 m |
Two consequences for rack planning:
- Switch-to-switch passive copper stops at 2 m. Not 3 m. The 3 m passive parts exist only in the splitter families. A cross-rack ToR-to-ToR run planned at 3 m passive does not have a part behind it.
- There is no 200G transceiver. 200G links are built by splitting a 400G transceiver across two fibres. Somebody will try to order a 200G optic for a 1:4 breakout; it is not a supply problem, the part does not exist.
Active Optical Cable / AOC
Multi-mode fiber with fixed active optical transceivers built into the connector heads. The transceiver is not pluggable -- it is permanently attached to the cable.
Maximum distance: 100m. Power consumption: 1-3W per cable. Cost: lower than pluggable SR4 for distances under 100m. Weight and bend radius: significantly lighter and more flexible than copper DAC at longer lengths.
Tradeoff: when the transceiver fails, the entire cable must be replaced. In a large cluster with thousands of AOC runs, this is manageable -- a field technician swaps the full assembly. In installations where the fiber run is difficult to access (through conduit, under raised floor), the inability to replace just the transceiver is a maintenance consideration.
Use for: leaf-to-spine connections in BasePOD (typically 10-30m). Also common for ToR-to-spine runs in end-of-row designs.
Pluggable SR4 / SR8
A pluggable OSFP or QSFP-DD transceiver module plus a separate OM4 MMF cable assembly with MPO connectors.
Maximum distance: 100m (SR4 and SR8 both at OM4). The transceiver and cable are independent components. Cost: higher than AOC, but transceiver and cable are replaced separately.
The key advantage: when a transceiver fails in a high-density deployment, you replace the 4kg transceiver, not the cable run through the raised floor. The fiber plant -- which may represent years of structured cabling investment -- stays in place.
Use for: any connection where the cable run has a lifetime significantly longer than the transceiver refresh cycle, or where transceivers are replaced frequently due to operational environment.
Pluggable DR4 / DR8 / FR4
Single-mode fiber transceivers for longer distances. Same pluggable model as SR4/SR8 -- separate transceiver and cable -- but using SMF instead of MMF.
DR4 (Data center Reach): 500m, MPO-12 connector. Cross-row and cross-hall. FR4 (Far Reach): 2km, LC duplex. Cross-building. LR4 (Long Reach): 10km, LC duplex. Campus-scale.
The selection decision tree
Same rack, <=3m: Passive DAC
Same rack, 3-7m (800G): AEC
Same row, 7-30m: AOC or pluggable SR4/SR8
Cross-row, 30-100m: Pluggable SR4/SR8 (OM4 MMF)
Cross-row, 100-500m: Pluggable DR4/DR8 (SMF)
Cross-building, <=2km: Pluggable FR4 (SMF)
Cross-building, <=10km: Pluggable LR4 (SMF)
Between DCs: ZR / DWDM coherent (not AI cluster fabric)