Cisco's USFP optic starts at 6.4 Tbps and reaches 12.8 Tbps at 32 lanes
Cisco, working with Nvidia, announced USFP, an OSFP-sized pluggable that targets 12.8 Tbps from 32 lanes of 400G. The first spec is 16 lanes at 6.4 Tbps, a 4x gain over 1.6 Tbps optics. Only the 32-lane version reaches 8x.
The Engineer · Build desk
Two figures for power per pluggable optic. A high-end pluggable today might pull on the order of 20 watts. Cisco's USFP spec calls for up to 100 watts per pluggable.
| Measure | Value | As of | Claim |
|---|---|---|---|
| High-end optic today, order of | 20 W | 9 | |
| USFP spec, up to | 100 W | 9 |
What happened
- Cisco's spec calls for up to 100 watts per module, where a high-end pluggable today draws on the order of 20 watts.
- Because of the added electrical contacts, Cisco says USFP modules will seat in two phases: insertion into the cage, then a second mechanism that makes the electrical connection.
- Nvidia's recent Spectrum-X and Quantum-X switches mostly use co-packaged optics, and dropping pluggables on the switch side saves roughly 30 percent of power.
- USFP will be an open standard that anyone can use in their designs, according to The Register.
Why it matters
- What it limits constraint Heat becomes the design limit. At the ceiling each port draws up to 5x the power of a 20 W part, so a fully populated faceplate has to be cooled and powered for that.
- Who's exposed exposure A single bad laser or DSP in a USFP module can take out whole systems, a wider failure domain than a conventional pluggable. Co-packaged optics are worse, with a rack's worth of accelerators lost on one failure.
- Who has to decide decision Clusters well below 100,000 GPUs, or racks with a top-of-rack switch, can keep lower-power direct attach copper. USFP is a question for large fabrics first.
The 4x-to-8x range The Register cites [5] is two different modules. Cisco's initial spec calls for 16 lanes of 400 Gbps, about 6.4 Tbps [6]. The 12.8 Tbps part needs all 32 lanes [2][3]. Against pluggables that cap near 1.6 Tbps on eight 200 Gbps links [4], those come to 4x and 8x [17]. The report does not include a ship date for either version or a power figure for the 16-lane one.
The cabling pitch splits the same way. The Register's example is eight GPUs, each paired with a 1.6 Tbps NIC, folded into one fiber bundle [7]. That takes the 12.8 Tbps module. On the switch side, a future 204.8 Tbps switch would need 16 USFP ports to serve 128 endpoints at 1.6 Tbps [8]. At 6.4 Tbps the same module covers four GPUs and the switch needs 32 ports [18].
Power per bit improves on paper. At the ceiling, a 12.8 Tbps module works out to about 7.8 W per Tbps, against 12.5 W per Tbps for a 20 W part at 1.6 Tbps [19]. The 100 W is a spec limit. The 20 W is an order-of-magnitude figure from the report for today's high-end parts [9].
Contact count is the mechanical cost. The report ties links to contacts one for one: four times as many links means four times as many contacts to seat [10]. Going from eight lanes to 32 is that 4x. On the same logic the 16-lane first spec carries 2x [20]. It gets its 4x bandwidth by doubling lane speed as well as lane count, so we'd expect it to face half the seating problem of the 32-lane part.
Nvidia is a co-developer of USFP [1]. The Register says it can only speculate on Nvidia's interest and guesses densification, with Nvidia moving toward 600 kW racks of 144 GPUs [21]. One number fits that guess. Vera Rubin Ultra promised about 3.2 to 6.4 Tbps of scale-out bandwidth per GPU package, depending on how it is counted [22], and the 16-lane spec is 6.4 Tbps [6].
In our view the 16-lane module is the one to plan around until a 32-lane date exists.
What to watch
- A ship date for the 16-lane module and a power figure for it, neither of which the report includes.
- A schedule for the 32-lane, 12.8 Tbps version, since only that version delivers the 8x gain.
- An Nvidia product that adopts USFP. The densification guess in the report stays untested until one appears.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption
- Insufficient
- Hype gap+20
- Incentives
- Insufficient
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Cisco developed the USFP pluggable in collaboration with Nvidia and announced it at the Open Compute Project global summit.
- [2]
USFP promises up to 32 lanes of 400 Gbps in a pluggable roughly the size of existing OSFP modules.
- [3]
Each USFP pluggable would have 12.8 Tbps of Ethernet bandwidth.
- [4]
The most sophisticated pluggable optics today cap out at around 1.6 Tbps using eight 200 Gbps links.
- [5]
The new pluggable will pack between four and eight times the bandwidth of existing pluggables into a single port.
- [6]
Cisco's initial USFP spec calls for 16 links of 400 Gbps per pluggable, or about 6.4 Tbps of bandwidth.
- [7]
A single USFP pluggable could consolidate eight GPUs' worth of cabling into a single fiber optic bundle, in datacenters where each GPU is paired with an 800 Gbps and soon 1.6 Tbps NIC.
- [8]
A future 204.8 Tbps switch would only require 16 USFP ports to support 128 endpoints at 1.6 Tbps.
- [9]
A high-end pluggable optic today might pull on the order of 20 watts; Cisco's USFP spec calls for up to 100 watts per pluggable.
- [10]
Four times as many links per pluggable means four times as many electrical contacts that have to be perfectly seated.
- [11]
Cisco says USFP modules will use a two-phase insertion mechanism: the pluggable is first inserted into the cage, then a second mechanism is actuated to make the electrical connection.
- [12]
USFP would have a bigger blast radius than conventional pluggables: a single bad laser or DSP could take out far more than a single GPU, potentially entire systems.
- [13]
In co-packaged or near-package optics, one failure could drop an entire rack's worth of accelerators from the network.
- [14]
Nvidia has largely moved away from pluggables in its most recent switches; many Spectrum-X and Quantum-X switches use co-packaged optics, and cutting out the pluggables on the switch side nets roughly 30 percent power savings.
- [15]
Nvidia's NICs still rely on pluggables in part because they give integrators more flexibility; not everyone is building 100,000-plus GPU clusters, and racks with a top-of-rack switch can use lower-power direct attach copper without much issue.
- [16]
USFP will be an open standard that anyone can use in their designs.
- [17]
The 16-lane 6.4 Tbps spec is 4x the 1.6 Tbps ceiling of today's pluggables, and the 32-lane 12.8 Tbps part is 8x, which accounts for the 4x-to-8x range.
- [18]
At 16 lanes (6.4 Tbps) one module covers four 1.6 Tbps GPU NICs and a 204.8 Tbps switch needs 32 ports; at 32 lanes (12.8 Tbps) it covers eight NICs and the switch needs 16 ports.
- [19]
At the 100 W ceiling a 12.8 Tbps module is about 7.8 W per Tbps, against 12.5 W per Tbps for a 20 W part at 1.6 Tbps; the module draws up to 5x the power for 8x the bandwidth.
- [20]
Against today's eight-lane pluggables, a 32-lane USFP has 4x the lanes and the 16-lane initial spec has 2x the lanes.
- [21]
The Register says it can only speculate on Nvidia's interest in USFP and that it likely comes down to densification, as Nvidia moves toward 600 kW racks packing 144 GPUs.
- [22]
Nvidia's Vera Rubin Ultra platform, first teased in early 2025, promised 115.2 TB/s of scale-out bandwidth per rack, or about 3.2-6.4 Tbps per GPU package depending on whether bidirectional bandwidth is counted.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- Cisco, Nvidia team up to build fatter, faster pluggable (4 minute read)
assets.theregister.com
1 article · October 15, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Pluggable opticsFollow
- AI Data Center NetworkingFollow
- Co-Packaged OpticsFollow