Nvidia and Cisco are collaborating on Ultra-Dense Small Form Factor Pluggable (USFP), a new form factor for optical modules with 400 Gb/s per lane. According to the two companies, USFP delivers four times the bandwidth of current pluggables with 200 Gb/s per lane. This enables companies to achieve up to 12.8 Tb/s in total capacity.
AI data centers need to connect an ever-increasing number of computing systems. This requires not only more bandwidth per connection but also more connections per switch. When a switch is full, you often need an additional network layer. This adds optical links and power and increases latency.
USFP is designed to solve this problem without eliminating the familiar pluggable transceiver. The first modules will feature 16 full-duplex lanes in a housing the size of an OSFP module. They are intended for both switches and SuperNICs. Later versions are expected to support 32 or more lanes. Because both the number of lanes and the per-lane speed double, AI networks can connect more 1.6 Tb/s endpoints without adding network layers. Total capacity per module reaches 12.8 Tb/s.
Compared to a comparable 200 Gb/s-per-lane module, USFP promises four times as much bandwidth and bandwidth density at the front of the switch. Energy efficiency is approximately 1.3 times higher.
New connector, no separate cooling
How do you fit so many lanes into such a small module? Nvidia and Cisco developed a new two-dimensional array connector for this purpose. It uses a two-step insertion method: the high-speed contacts don’t make contact until after the module is installed in the system. This reduces contact wear, improves signal integrity, and makes the required insertion force more predictable. The connector is designed for 400 Gb/s, but according to the companies, it also delivers better performance at 200 Gb/s per lane.
In addition, a USFP module can dissipate up to 100 watts of heat without a separate built-in cooling circuit. This should simplify switch design. All configurations also support optical designs that simplify signal processing, which should reduce latency and power consumption.
Bill Gartner, who is responsible for optical systems at Cisco, explains why a new form factor was chosen: “Cisco and Nvidia recognized that scaling pluggable optics to 400Gb/s per lane would require more than another incremental extension of today’s module designs.”
TE Connectivity and Amphenol are also collaborating on the hardware. They will manufacture cages, connectors, and modules, and more partners are expected to join. The companies have not yet announced when the first USFP modules will become available.
Tip: Cisco and Nvidia lower the barrier to secure, full-stack AI infrastructure