SR4 vs. DAC Cabling Guide

→ View the MMA1T00-HS 200G QSFP56 HDR SR4 Transceiver

SR4 Optical Transceivers vs. DAC Cables for HDR InfiniBand: How to Choose

If you're wiring a 200Gb/s HDR InfiniBand fabric — a DGX A100 node, an HGX build, or any Mellanox/NVIDIA ConnectX-6 deployment — you'll hit this decision at every single link: copper DAC, active optical AOC, or an SR4 transceiver on structured MPO fiber?

All three show up as QSFP56 ports. All three will pass HDR traffic. But they are not interchangeable, and picking the wrong one for a given run either wastes money or leaves you unable to reach the port you need. Here's the short version, then the reasoning behind it.

The Bottom Line

• Under 5 meters, same rack or adjacent rack: use a passive DAC cable. It's the cheapest option, draws no additional power, and has the lowest latency because there's no optical serialization.
• 5–100 meters, fixed point-to-point run, no repatching planned: an AOC cable does the job in one part number — fiber reach with DAC-style plug-and-play.
• 5–100 meters where the fiber path is shared infrastructure — cross-row, into a centralized fabric rack, or anywhere you expect to reconfigure or extend later — use an SR4 transceiver, like the MMA1T00-HS, on OM4 MPO trunk cabling run through a patch panel.

The first two calls are simple. The third is the one worth explaining, because on paper an AOC and an SR4 transceiver both max out around 100 meters over the same multimode fiber — so why would you ever choose the transceiver?

DAC vs. AOC vs. SR4 Transceiver, Side by Side

The reach numbers for AOC and SR4 optics are genuinely the same — both are built around the same 100m OM4 multimode limit. The difference is architectural, not electrical.

Why Choose a Transceiver Over an AOC When the Reach Is Identical

An AOC is a sealed, fixed-length assembly. The fiber and the optics on both ends are one part. If a single strand degrades, or you need to move one end six feet to a different port, you replace the entire cable.

An SR4 transceiver is a module that plugs into a QSFP56 port and terminates into an MPO-12 connector, which then runs into your own structured multimode fiber — trunk cables, patch panels, cross-connects. That gets you three things AOC can't:

• Independent failure domains — if one side of a link goes bad, you swap a single module, not a cable that happens to be 80 meters long.
• A reusable fiber plant — the MPO trunk cabling between a compute row and a fabric rack gets installed once. As switches get added, moved, or upgraded to NDR, you're re-terminating or repatching, not pulling new cable through cable trays.
• Mixed reach on one physical run — a structured OM4 plant patched through an intermediate distribution area can serve multiple logical link lengths from the same cabling, which a fixed AOC cannot.

This is exactly why hyperscale and reference cluster designs treat structured fiber with pluggable optics as the default for anything beyond a single rack — NVIDIA's own DGX SuperPOD reference architecture separates GPU compute racks from centralized fabric and management racks connected by structured cabling, rather than point-to-point AOC runs everywhere.

Where This Shows Up in a Real HDR Fabric

NVIDIA's DGX A100 SuperPOD reference design is the clearest example. Each DGX A100 node carries eight HDR 200Gb/s links — one per GPU, through dual-port ConnectX-6 adapters — in a 1:1 GPU-to-NIC ratio. Those links land on Mellanox Quantum QM8700 HDR switches arranged in a non-blocking fat-tree topology, with leaf switches serving compute racks and spine switches typically centralized in dedicated fabric racks.

That's precisely the layout where the DAC/AOC/transceiver decision plays out link by link:

• DGX node to leaf switch, same rack: DAC, if the leaf lives in the compute rack.
• Leaf to spine, adjacent racks: AOC works fine if the topology is fixed and won't be repatched.
• Leaf or spine into a centralized fabric row, or any link that crosses into shared structured cabling: SR4 transceivers on MPO trunk fiber, so the fabric can be re-patched as the pod grows without re-running point-to-point cables.

At pod scale — hundreds of links — the cost, spares, and reconfiguration math consistently favors structured fiber with pluggable optics over a wall of fixed-length AOCs.

A Note on MPO Connectivity

SR4 modules like the MMA1T00-HS use an MPO-12 UPC connector, but only 8 of the 12 fiber positions carry traffic — four for transmit, four for receive. When you're speccing trunk cabling and patch panels, keep polarity method (Type A/B/C) consistent across your fiber plant; mismatched polarity between trunks, cassettes, and modules is one of the most common causes of a "link down" that isn't actually a bad transceiver.

Compatible vs. OEM Optics

Whichever direction you land on, the transceiver itself doesn't need to be a genuine Mellanox/NVIDIA part to work correctly. A properly EEPROM-programmed compatible module — like the MMA1T00-HS — is read by ConnectX-6 adapters and Quantum switches as a native module: same 850nm wavelength, same 100m/70m reach, same DDM telemetry, typically at a significant discount to OEM pricing. The things worth checking before you buy compatible optics are the same three things regardless of vendor: correct EEPROM coding for your specific platform, DOM/DDM support so you can monitor link health, and a warranty that actually covers the module if it fails in service.

Frequently Asked Questions

Can I mix DAC, AOC, and transceiver links in the same fabric?
Yes — this is normal. Most real deployments use all three, matched to distance and topology link by link, as shown above.

Does an SR4 transceiver cost more than an AOC for the same run?
Per link, often yes once you count both modules plus fiber. At fabric scale, where the fiber plant is shared across many links and gets reused as the topology changes, structured cabling is typically the cheaper long-term choice.

Will a compatible SR4 transceiver trigger "unsupported optics" warnings on a Mellanox/NVIDIA switch?
Not if it's properly EEPROM-programmed for that platform. That's the whole point of buying a tested-compatible module rather than a generic one.

What reach do I actually get with the MMA1T00-HS?
100 meters over OM4 multimode fiber, 70 meters over OM3, per the NVIDIA HDR QSFP56 MMF transceiver specification.

The Resilient Tec Guarantee

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

Need help mapping DAC, AOC, and transceiver links across your HDR InfiniBand topology before you order? Contact Resilient Tec — we'll review your rack layout and confirm compatibility before anything ships.

See also our DAC Cables Guide, AOC Cables Guide, and Optical Transceivers Guide.