OSFP, The 800G workhorse, eight lanes of 100G PAM4 for AI/ML and hyperscale.
Octal Small Form-factor Pluggable, OSFP MSA. The primary 800G transceiver form factor, eight electrical lanes at 100G PAM4 each. Physically wider than QSFP-DD; not mechanically interchangeable. Used in AI/ML training clusters (NVIDIA Quantum-2 InfiniBand, Spectrum-X Ethernet), hyperscale spines, and 800G coherent DCI. CMIS 5.x management.
Specifications
MSA references
OSFP MSA rev 5.x (mechanical) · CMIS 5.0 / 5.1 (management) · IEEE 802.3df / 802.3dj (800G PHY)
Electrical interface
8 lanes × 53.125 Gbaud PAM4 (100G per lane)
Power consumption
15 W – 20 W typical; up to 23 W for coherent ZR variants
Signalling
PAM4 with KP4 / RS-FEC
FEC
RS-FEC (544,514) required; coherent ZR uses OpenZR+ MSA FEC layer
Connector types
MPO-12 (DR8 breakout to MPO-16) · MPO-16 (SR8, DR8) · LC duplex (LR8, ER8, ZR)
Available SKUs
| Part number | Spec | Wavelength | Fiber | Reach | Notes |
|---|---|---|---|---|---|
NAP-OSFP-SR8 | 800GBASE-SR8 | 850 nm | MM OM4 / OM5 | 100 m (OM4) / 150 m (OM5) | MPO-16, 8 parallel |
NAP-OSFP-DR8 | 800GBASE-DR8 | 1310 nm | SM parallel | 500 m | MPO-16 |
NAP-OSFP-LR8 | 800GBASE-LR8 | LAN-WDM 8λ | SM | 10 km | LC duplex |
NAP-OSFP-ER8 | 800GBASE-ER8 | LAN-WDM 8λ | SM | 40 km | Extended reach single-mode |
NAP-OSFP-800ZR | 800G coherent ZR | C-band | SM | 80–120 km | DSP-based, no amplifier |
NAP-OSFP-800ZR-PLUS | 800G ZR+ (OpenZR+) | C-band | SM | 500 km+ with amps | Carrier-grade DCI |
NAP-OSFP-DAC-1M…2M | 800G DAC | — | Twinax | 1–2 m | Passive intra-rack |
NAP-OSFP-AOC-3M…30M | 800G AOC | 850 nm | — | 3–30 m | Dominant in AI clusters |
NAP-OSFP-BREAKOUT | 800G → 2×400G DR4 | 1310 nm | SM parallel | 500 m | Standard fabric breakout |
In production with
OSFP modules are deployed across these platforms today. Full coding details are on the compatibility matrix.
- NVIDIA Quantum-2 InfiniBand (NDR 400G / NDR 800G)
- NVIDIA Spectrum-4 SN5000 series
- Arista 7060X6, 7800R3 800G line cards
- Cisco Silicon One 800G platforms
- Juniper PTX10000 series
- Hyperscale custom silicon (Meta, Google, ByteDance)
Operational notes
Recommended for
- NVIDIA Quantum-2 / Spectrum-X AI training fabric
- Hyperscale 800G spines (Meta, Google, ByteDance-class)
- Disaggregated NVMe-oF storage at 800G
- Long-haul DCI with coherent 800ZR / ZR+