AIbaseUpdated

NVIDIA and Cisco team up to unveil the USFP optical module spec: 400G per lane, four times the bandwidth, giving AI clusters' blood vessels…

AI compute is often bottlenecked by data transfer between chips. On October 12, NVIDIA and Cisco launched USFP, the industry's first pluggable optical module spec supporting 400Gbps/lane high-speed signaling, to widen…

When AI compute is pushed to its limits, it often gets stuck on the data-carrying "blood vessel" between chips. On October 12 local time, NVIDIA and Cisco jointly launched USFP, the industry's first pluggable optical module spec supporting 400Gbps/lane high-speed signaling, effectively widening this bottleneck's channels by a significant margin.

Compared with existing OSFP optical modules, USFP is similar in size but packs a denser core. The first-generation products integrate 16 full-duplex lanes, and can later expand to 32 or more, with throughput topping out at 12.8Tb/s. The key multiplier is stated plainly: its bandwidth reaches 4 times that of current 200G/lane pluggable optical modules; backed by 400G SerDes technology, its energy efficiency is also 1.3 times higher than the latter.

The connector has been redesigned this time. The new two-dimensional array connector uses a two-step insertion scheme—the module is inserted first, and the high-speed contacts engage only in the second step, which reduces contact wear, keeps signal integrity stable, and makes the force required for each insertion more predictable. It supports 100W power consumption, which in turn simplifies the switch-side design, with no extra hassle over power delivery and cooling.

For NVIDIA, the significance of this move lies not in the module itself, but in the network it wants to defend. As large model training clusters spread ever wider, thousands of accelerator cards must exchange massive gradients and activations in real time, and the bandwidth and energy efficiency of optical interconnects directly determine whether the whole cluster runs smoothly or its components idle waiting on each other. USFP doubles the per-lane rate and quadruples total bandwidth, effectively laying thicker data pipelines in advance for the next generation of larger-scale AI factories. When the chip giant and the veteran network equipment maker put this spec on the table together, the most overlooked yet most critical stretch of "neural wiring" in the AI race has already begun its generational upgrade ahead of schedule.

Original source

AIbase

Content notes

Original publication and rights belong to the source.

Machine translation · Refer to the original