甚麼是 USFP？面向 AI 時代的超高密度小型可插拔光模組

AI training clusters are running out of three things at once: rack space, power, and fiber. A modern GPU cluster interconnects tens of thousands of accelerators, and the scale-out fabric that ties them together is built from pluggable optics. When every port is an OSFP or QSFP-DD module, a 1U switch tops out around 32 × 800G — and each module can draw 15–30 W. Multiply that across a 102.4T switch and the optics alone consume kilowatts, while the faceplate runs out of room and the fiber plant becomes a spaghetti wall of MPO trunks.

USFP attacks this directly. By shrinking the module to roughly half the width, USFP doubles faceplate density — up to ~64 × 800G per 1U — while targeting lower per-module power. The payoff is more fiber density and cable density per rack, fewer switches to reach the same GPU count, and a better watts per terabit figure for the whole AI data center.

What Exactly Is USFP?
USFP (Ultra Small / Ultra-Dense Small Form-factor Pluggable) is a very compact optical transceiver form factor defined by an industry MSA. Like OSFP and QSFP-DD, it is a hot-pluggable module that sits in a cage on the switch faceplate — but the cage and module are markedly narrower. USFP uses a short electrical reach to the host ASIC and is specified around 100G and 200G electrical lanes, carrying PAM4 signaling. An 800G USFP module aggregates eight 100G lanes or four 200G lanes, and the same mechanical platform is intended to scale toward 1.6T as 200G per lane matures (aligned with IEEE 802.3dj and OIF 224G work).

The defining idea is simple: more ports, less power, less fiber per bit. USFP does not invent new physics — it re-packages proven silicon photonics, lasers, and DSP/LPO electronics into a denser shell so that density, not raw reach, is the win.

Challenges and Caveats
Young ecosystem: fewer module and switch options than OSFP/QSFP-DD; validate interoperability.
Reach trade-off: density-optimized, so long-haul reach is not its game — keep OSFP/QSFP-DD for ZR/long DCI.
Thermals at the faceplate: more ports per RU concentrates heat; confirm switch airflow and cooling.
Fiber plant: higher density changes MPO/trunk counts and breakout planning — design the cabling, not just the switch.
How to Choose
Pick USFP when port density and power are your binding constraints — AI scale-out, dense leaf tiers, and short-reach DCI. Stay with OSFP/QSFP-DD for long reach, multi-vendor fabrics, and mature ecosystems. If you are power-limited at the rack edge, evaluate USFP + LPO, and pilot CPO where density justifies it.

Frequently Asked Questions
What does USFP stand for?
USFP stands for Ultra Small Form-factor Pluggable (often described as Ultra-Dense Small Form-factor Pluggable) — a compact pluggable optical transceiver form factor defined by an industry MSA for high-density, lower-power links.

How is USFP different from OSFP and QSFP-DD?
USFP is roughly half the width, doubling faceplate density (up to ~64×800G per 1U vs ~32) and targeting lower power — optimized for density and AI scale-out. OSFP and QSFP-DD are wider, higher-power, and stronger for long reach and multi-vendor interoperability.

Is USFP 800G or 1.6T?
USFP is shipping for 800G today (8×100G or 4×200G PAM4 lanes), with 1.6T the roadmap target as 200G-per-lane optics and SerDes mature.

How does USFP relate to CPO and LPO?
USFP is the pluggable bridge: its compact, low-power design is LPO-friendly today and delivers CPO-like density while remaining field-replaceable — a stepping stone toward near-packaged/co-packaged optics and optical I/O.

Is USFP a replacement for OSFP?
No — they coexist. Use USFP where density and watts dominate (AI scale-out, dense leaf); use OSFP/QSFP-DD for long reach, ZR/DCI, and where you need the broadest multi-vendor ecosystem.
More on : The Network DNA