Introduction: Why Cabling Decides AI Fabric Performance
You can spend millions on NVIDIA Quantum-2 NDR 400G InfiniBand switches and ConnectX-7 adapters, but if the cable connecting them is wrong—wrong reach, wrong connector, wrong BER tolerance—the fabric will drop packets, throttle adaptive routing, and silently degrade your all-reduce latency by 30 percent or more. In 2026 AI clusters, cabling is not an afterthought. It is the physical layer that every RDMA transaction, every GPUDirect RDMA transfer, and every SHARP aggregation traverses.
The cable is the smallest, cheapest component in the rack—until it fails. Then it becomes the most expensive one.
This guide breaks down the three dominant cabling technologies in modern data centers—Direct Attach Copper (DAC), Active Optical Cable (AOC), and optical transceivers (SR4/DR4/FR4/LR4)—and maps each to the right use case for 100G, 200G, 400G, and 800G networks. Whether you are building a 64-node InfiniBand NDR fabric or a 1,024-GPU Ethernet RoCE cluster, the decisions below will determine your per-port cost, power budget, thermal load, and mean time between failures (MTBF).
The Three Cabling Families at a Glance
| Attribute | DAC (Twinax Copper) | AOC (Active Optical) | Optical Transceiver + Fiber |
|---|---|---|---|
| Reach | 0.5 m – 5 m | 1 m – 100 m | 100 m – 40 km |
| Latency | ~0.1 ns/m (lowest) | ~5 ns (DSP adds delay) | ~5 ns + 5 ns/m fiber |
| Power per port | 0 W (passive) / <1 W (active) | 0.5 – 1.5 W | 1.5 – 12 W (speed-dependent) |
| Cost (400G, 3 m link) | Quoted | Quoted | varies by configuration and volume (transceiver only) |
| Bend radius | ~25 mm (rigid) | ~10 mm (flexible) | ~10 mm (patch cord) |
| Weight (400G, 3 m) | ~220 g | ~90 g | ~60 g (transceiver) + 40 g (fiber) |
| Connector | QSFP-DD / OSFP / SFP-DD | QSFP-DD / OSFP / SFP-DD | QSFP-DD / OSFP (module) + MPO/LC (fiber) |
| Thermal output | Negligible | ~0.5 W per end | Up to 12 W per module |
The key insight: DAC wins on power and latency for short reaches inside a rack; AOC bridges the gap for intra-row links; and optical transceivers are the only option for inter-row, inter-rack, and data-center interconnect (DCI) distances.
DAC: The Workhorse of Intra-Rack Links
Passive DAC
Passive DAC cables are twinaxial copper pairs with no active electronics. The switch’s PHY drives the signal directly. At 25G per lane (100G QSFP28), passive DAC works up to 5 meters. At 50G per lane (200G/400G QSFP56/QSFP-DD), the practical limit drops to 2–3 meters due to PAM4 signal integrity constraints. At 100G per lane (800G OSFP/QSFP-DD800), passive DAC barely reaches 1 meter.
| Speed | Connector | Lanes × Rate | Max Passive Reach | Typical Use Case |
|---|---|---|---|---|
| 25G | SFP28 | 1 × 25G | 5 m | Top-of-rack to server |
| 100G | QSFP28 | 4 × 25G | 3 m (3m) / 5 m (lower BER) | Leaf-to-spine intra-row |
| 200G | QSFP56 | 4 × 50G PAM4 | 2 m | Switch-to-switch intra-rack |
| 400G | QSFP-DD / OSFP | 8 × 50G PAM4 | 1.5 – 2 m | Spine-to-leaf adjacent rack |
| 800G | OSFP / QSFP-DD800 | 8 × 100G PAM4 | 0.5 – 1 m | Backplane / adjacent switch |
Active DAC
Active DAC cables embed a redriver or retimer IC at one or both ends to extend reach and improve signal integrity. They consume <1 W per port and can push 400G to 3–4 meters and 800G to 2 meters. The trade-off is slightly higher cost (an amount quoted per part numbermore than passive) and a small latency penalty (~2 ns). For dense 800G spine-to-leaf links across a single rack row, active DAC is often the sweet spot.
When to Choose DAC
- Intra-rack switch-to-server links at 25G/100G (TOR model)
- Spine-to-leaf in adjacent racks (≤ 2 m at 400G)
- Power-constrained deployments where every watt matters (passive DAC = 0 W)
- Ultra-low-latency HPC workloads where sub-nanosecond link delay matters
AOC: The Flexible Middle Ground
Active Optical Cables embed a VCSEL or silicon photonics transmitter and a PIN photodiode receiver inside the connector shell, terminated by a fixed-length fiber bundle. Unlike discrete transceivers, AOCs are pre-terminated, fully tested end-to-end, and cannot be field-reconfigured.
VCSEL-Based AOC (MMF, OM3/OM4)
For 100G and 200G, VCSEL-based AOCs on multimode fiber dominate. They are cheap (varies by configuration and volumefor 100G QSFP28), consume ~0.5 W per end, and support reaches up to 70 m (OM4) or 100 m (OM5). However, VCSEL technology struggles above 50G/lane. For 400G, most VCSEL AOCs use 8 × 50G PAM4, but reach drops to 30–50 m on OM4.
Silicon Photonics AOC (SMF)
For 400G and 800G, silicon photonics AOCs on single-mode fiber are the go-to. They use CW lasers and external modulators to achieve 500 m to 2 km reach. Cost is higher (varies by configuration and volumefor 400G), but power is still reasonable (1–1.5 W per end). The key advantage over discrete transceivers is that the optical path is sealed, reducing connector contamination issues.
| AOC Type | Speed | Fiber | Max Reach | Power/End | Cost (3 m) |
|---|---|---|---|---|---|
| QSFP28 100G AOC | 100G | OM4 MMF | 70 m | 0.5 W | Quoted |
| QSFP56 200G AOC | 200G | OM4 MMF | 50 m | 0.7 W | Quoted |
| QSFP-DD 400G AOC (VCSEL) | 400G | OM4 MMF | 30 m | 1.0 W | Quoted |
| QSFP-DD 400G AOC (SiPh) | 400G | SMF | 500 m | 1.5 W | Quoted |
| OSFP 800G AOC (SiPh) | 800G | SMF | 300 m | 2.0 W | Quoted |
When to Choose AOC
- Intra-row links spanning 5–30 m where DAC cannot reach
- Cable management in high-density racks where copper stiffness is a problem (AOC is 60% lighter and far more flexible)
- Airflow optimization — AOC’s smaller bend radius doesn’t block hot-aisle exhaust
- Pre-tested reliability — sealed optical path eliminates end-face contamination at one end
Optical Transceivers: The Scalable Backbone
For inter-row, inter-pod, and DCI links, pluggable optical transceivers are the only option. They decouple the optical fiber from the transceiver, allowing field replacement and reach customization. The trade-off is cost (varies by configuration and volumeper 400G module), power (up to 12 W for 800G), and the need to manage fiber connector cleanliness.
The 400G/800G Transceiver Landscape
| Module | Speed | Fiber | Reach | Power | Typical Price | Use Case |
|---|---|---|---|---|---|---|
| QSFP-DD 400G SR4 | 400G | OM4/OM5 MMF | 70–100 m | 7 W | Quoted | Intra-pod, intra-row |
| QSFP-DD 400G DR4 | 400G | SMF | 500 m | 7 W | Quoted | Inter-row, inter-pod |
| QSFP-DD 400G FR4 | 400G | SMF (duplex) | 2 km | 7 W | Quoted | Inter-pod, campus |
| QSFP-DD 400G LR4 | 400G | SMF (duplex) | 10 km | 8 W | Quoted | DCI, metro |
| QSFP-DD 400G ZR/ZR+ | 400G | SMF (duplex) | 80–120 km | 12 W | Quoted | Long-haul DCI |
| OSFP 800G DR8 | 800G | SMF | 500 m | 12 W | Quoted | Spine-to-spine, AI fabric |
| OSFP 800G 2x400G FR4 | 800G | SMF (duplex ×2) | 2 km | 10 W | Quoted | 800G breakout to 400G |
SR4 vs DR4 vs FR4: What’s the Difference?
- SR4 (Short Range, 4-lane): Uses VCSEL transmitters on multimode fiber (OM4/OM5). Cheapest option (varies by configuration and volumefor 400G), but limited to 70–100 m. Best for intra-pod links within the same row.
- DR4 (Datacenter Range, 4-lane): Uses silicon photonics or DML on single-mode fiber. Reaches 500 m. Moderate cost (varies by configuration and volumefor 400G). The default choice for inter-row links in most AI clusters.
- FR4 (Far Range, 4-lane): Uses 4 CWDM wavelengths on a single duplex SMF pair. Reaches 2 km. Slightly more expensive than DR4 but uses only 2 fibers (vs 8 for DR4), saving fiber infrastructure cost in large deployments.
- LR4 (Long Range, 4-lane): Similar to FR4 but reaches 10 km. Used for DCI between nearby data center buildings.
- ZR/ZR+ (Zebedo Range): Coherent optics reaching 80–120 km. Used for metro DCI. Power-hungry (12 W) and expensive (varies by configuration and volume).
Power and Thermal Implications
In a 128-port 400G QSFP-DD spine switch, if every port uses a DR4 transceiver drawing 7 W, the optical modules alone consume 896 W—nearly 40% of a 2,200 W switch budget. If you use passive DAC for intra-rack links instead, that drops to near zero for those ports. This is why most AI cluster designs use a hybrid approach:
| Link Type | Distance | Recommended Cabling | Power/Port | % of Fabric Links |
|---|---|---|---|---|
| Server-to-TOR | 1–3 m | Passive DAC (25G/100G) | 0 W | ~40% |
| TOR-to-Leaf | 3–7 m | Active DAC or AOC (100G/200G) | 0.5–1 W | ~25% |
| Leaf-to-Spine | 7–30 m | AOC (200G/400G) | 1–1.5 W | ~20% |
| Spine-to-Spine | 30–500 m | Transceiver DR4 (400G/800G) | 7–12 W | ~10% |
| DCI | >500 m | Transceiver FR4/LR4/ZR | 7–12 W | ~5% |
This hybrid strategy reduces total optical power by 50–60% compared to an all-transceiver design, which directly translates to lower cooling load and lower PUE.
Reliability and Field Failure Modes
In production AI fabrics, cable failures are the #1 cause of silent packet drops. A single degraded AOC can reduce NCCL all-reduce throughput by 15% without triggering any switch-level alarm.
| Failure Mode | DAC | AOC | Transceiver | Detection Method |
|---|---|---|---|---|
| Connector contamination | Low (copper contact) | Medium (optical end-face) | High (fiber end-face) | DOM TX/RX power |
| Bend-induced attenuation | N/A (copper) | Low (MMF tolerant) | Medium (SMF sensitive) | RX power threshold |
| Thermal degradation | N/A | Low (VCSEL drift) | Medium (laser aging) | DOM temperature |
| EMI / crosstalk | Medium (unshielded) | None (optical) | None (optical) | BER counter |
| Physical damage (crush, kink) | Low (robust) | Medium (fiber break) | Medium (fiber break) | Link down alarm |
| DSP failure | N/A (passive) | Low (single IC) | Medium (complex DSP) | Module reset count |
Best Practices for Cable Reliability
- Use Digital Optical Monitoring (DOM): All modern transceivers support DOM. Monitor TX power, RX power, temperature, and bias current. Set thresholds at 3 dB below nominal RX power to catch degradation before it causes packet loss.
- Clean fiber connectors: 80% of optical link failures are caused by contaminated end-faces. Use a cassette cleaner (not wipes) before every insertion. Inspect with a fiberscope if RX power is below threshold.
- Label both ends: In a 400G fabric with 512 cables, unlabeled links cost hours of tracing time during incidents. Use color-coded labels by speed (e.g. blue = 100G, yellow = 400G, red = 800G).
- Plan airflow around cable bundles: A bundle of 48 AOCs crossing a hot aisle can raise cable temperature by 10°C, shortening VCSEL lifetime by 50%. Route cables in cold-aisle cable trays wherever possible.
- Burn-test before deployment: Run 48-hour BER tests at line rate on every AOC and transceiver before racking. Field replacement of a buried spine link costs 10× the cost of pre-deployment testing.
Cost Optimization: New vs Used Cabling
Cabling is one of the largest line items in a network fabric budget. For a 1,024-GPU cluster with 3,072 400G links, transceiver costs alone can reach varies by configuration and volume. Used cabling—particularly original-vendor-grade DAC and AOC from Mellanox/NVIDIA, Finisar, and Amphenol—can reduce this by 40–60% with no measurable performance difference.
| Component | New price | Indicative used price | Savings | Warranty |
|---|---|---|---|---|
| 100G QSFP28 DAC (3 m) | Quoted | Quoted | 58% | 1 year |
| 100G QSFP28 AOC (5 m) | Quoted | Quoted | 60% | 1 year |
| 400G QSFP-DD DR4 | Quoted | Quoted | 59% | 1 year |
| 400G QSFP-DD SR4 | Quoted | Quoted | 57% | 1 year |
| 800G OSFP DR8 | Quoted | Quoted | 51% | 1 year |
At Zhongxin Data, Models listed on zhongxindata.hk are available original-vendor cabling for NVIDIA/Mellanox, Broadcom, and Cisco platforms. Every cable is burn-tested at line rate, DOM-verified, and shipped with a 1-year replacement warranty.
Conclusion: A Hybrid Strategy Wins
No single cabling technology is optimal for every link in an AI fabric. The most cost-effective and reliable designs use a layered approach: passive DAC for intra-rack, AOC for intra-row, and optical transceivers only for inter-row and DCI links. This strategy minimizes power consumption (critical for liquid-cooled deployments), reduces connector contamination risk, and cuts total cabling cost by 30–50% versus an all-transceiver approach.
At Zhongxin Data, we help data center architects specify the right cabling mix for their fabric topology. Browse our InfiniBand switches, Ethernet switches, and full catalog of network adapters, cables, and transceivers. For bulk pricing and fabric design consultation, contact us on WhatsApp: +86 13068616914.
Full specs and current availability are on zhongxindata.hk.
Which cabling family belongs in each part of the fabric
- DAC inside the rack for short reaches where power and cost dominate; AOC where the run is longer but still in-rack; optics for anything crossing rows or halls
- The reach limit is a hard physical constraint — a passive DAC cable has a maximum length beyond which it will not link reliably at the rated speed
- The switch port type constrains the connector, so the port decision and the cabling decision have to be made together
Where our switch and adapter families land on cabling
- QSFP28 ports take 100GbE optics and DAC; QSFP56 ports take 200GbE; OSFP ports take 800GbE and are the forward-looking choice
- We supply the switch and adapter families these cables plug into, and availability is confirmed per part number
- Send the port type, the reach and the quantity and we quote per part number
How we verify optics and cabling on the hardware we list
- Every unit is cleaned, re-flashed and bench-tested on our own test line before it is listed
- For switches and adapters we confirm negotiated link rate, firmware revision and error counters against the part number you send
- Ask us for photos of the actual unit before you commit
Related reading
- QSFP28 vs QSFP-DD vs OSFP — the connector decision that precedes cabling
- liquid cooling — the thermal constraint on in-rack cabling
- spine-leaf architecture — which runs actually cross rows
- network switches — the port types these cables plug into
Frequently Asked Questions
Can I mix DAC and AOC on the same switch port?
Yes. A switch port does not care whether the module is passive copper, active copper or optical — what matters is that the module matches the port’s form factor and speed, and that the module’s reach covers the run. Mixing media across ports in the same fabric is normal; mixing speeds and expecting identical error margins is not.
What is the difference between OM4 and OM5 multimode fibre, and does it matter for 400G?
OM4 is rated to 400Gb/s over roughly 100 metres while OM5 extends that to around 400 metres at the same speed. For 400G the practical difference is that OM4 forces a shorter reach, which pushes more runs onto active copper or multimode with fewer connectors. If your runs exceed the OM4 limit, plan the optics type before the cabling route.
How do I calculate the power budget for optical modules in my switch?
Multiply the module’s power draw by the number of populated ports and compare against the switch’s optical power budget and its total port power. At 800G a single module can draw 20-30W, so a fully populated 64-port switch can add more than 1.5kW of optics load on top of the switch itself — which changes the rack power and cooling design.
Are third-party transceivers compatible with NVIDIA/Mellanox switches?
Generally yes, because the form factor and electrical interface are standard, but compatibility is not the same as feature parity. Vendor firmware may expose extra diagnostics, and a module without full digital diagnostic monitoring will report less than an original module. For a fabric you intend to support, test the module in the switch before committing to a volume order.
What is the typical MTBF for DAC, AOC and optical transceivers?
Passive DAC has the fewest components and typically the highest MTBF. Active copper adds retimers and power, which is where failures concentrate. Optical transceivers have the most components — laser, driver, thermal management — and their practical lifetime is often shortened by heat in a badly ventilated rack as much as by component ageing.