InfiniBand optics & cables
Đội ngũ của chúng tôi đang thực hiện bản dịch. Bài viết này tạm thời chưa có bản tiếng Việt và đang được hiển thị bằng tiếng Anh.
InfiniBand uses the same pluggable housings as high-speed Ethernet, but the choice of cable versus transceiver — and the tolerance for third-party modules — follows its own rules shaped by HPC and AI clusters.
Cable or transceiver, by generation
| Generation | Port | Copper DAC | Active optical cable | Transceiver + fibre |
|---|---|---|---|---|
| EDR 100G | QSFP28 | up to ~3–5 m | up to 100 m | SR4 (MPO-12 MMF), LR4 |
| HDR 200G | QSFP56 | up to ~2 m | up to 100 m | SR4/DR4 |
| NDR 400G | OSFP (twin-port 2×400G) | ~1–2 m | limited | DR4/DR8 on MPO-12 APC single-mode |
The pattern: as rates rise, passive copper reach collapses and fabrics move to optics. At NDR the norm is single-mode DR optics with MPO-12 APC connectors — the angled polish is mandatory, and mixing APC with flat UPC connectors is a common installation fault.
Twin-port OSFP
NDR switch ports are commonly twin-port OSFP: one OSFP module carries two independent 400G links (2×DR4), doubling faceplate density. A finned integrated heat sink handles the power. See OSFP.
Signalling and FEC
- EDR: NRZ 25G lanes; HDR: PAM4 50G; NDR: PAM4 100G.
- RS-FEC is mandatory from HDR onward. FEC mismatch or a module that does not advertise the expected FEC is a frequent "link never trains" cause.
Vendor policy
InfiniBand fabrics are dominated by a single switch/NIC vendor, and that vendor's platforms validate cables and optics by identity. A module that is not recognised may read correctly yet fail to link with a status such as "speed and type: Not supported" — see Vendor lock. Operators therefore stock qualified optics and, where third-party modules are used, code them to the expected identity.
Operating at scale
- Thousands of optics per cluster: even a small annual failure rate means a steady stream of replacements. Track per-lane DDM to catch a dying lane before it slows a collective operation.
- Cleanliness on MPO end-faces is the dominant cause of flaps — see CMIS issues for 400G specifics.
In CodingBox
CodingBox reads EDR/HDR modules per SFF-8636 and NDR OSFP/QSFP-DD per CMIS: identity, module state, per-lane DDM, and the identity fields that determine whether a fabric will accept the module.