In layman’s terms: New data center projects are being built everywhere, with no end in sight. Traditional connections between data centers are becoming increasingly inefficient. That’s why Nvidia and Cisco have developed a new standard that aims to transfer all those terabytes of data as quickly and efficiently as possible.
GPU manufacturer Nvidia and network specialist Cisco have introduced a new network interface called Ultra-Dense Small Form Factor Pluggable (USFP). The new optical transceiver is intended to significantly increase bandwidth in data centers and help manage the enormous data requirements of modern AI models and training workloads.
According to Cisco, the USFP module is the first interface to support 400 Gbps per lane. Ultimately, the module should support total data transfer rates of up to 12.8 Tbit/s. USFP’s total bandwidth capacity is four times that of optical interfaces at 200 Gbps per lane, while providing a 1.3x improvement in energy efficiency.
According to Cisco, USFP is designed to give data center customers access to next-generation optical networking technology that can accommodate the growth of their AI projects. The new interface is intended to scale efficiently using an established deployment model, although several development steps are still required before it reaches the full 12.8 Tbit/s capacity envisioned by the designers.
The first USFP-based optical transceivers are expected to feature 16 full-duplex lanes in a small form factor octagonal package. These pluggable interfaces support both switches and SuperNIC network adapters. The latter are special accelerator cards designed by Nvidia and its partners to transfer data directly between GPUs using the RDMA over Converged Ethernet (RoCE) network protocol. In simple terms, RoCE enables direct remote access to storage via Ethernet networks.
The next phase will expand the number of lanes to 32 or more, potentially allowing AI network systems to connect additional 1.6 Tbps endpoints without adding additional network layers. Cisco expects the complete USFP module to ultimately achieve a total bandwidth capacity of 12.8 Tbit/s.
The company also highlights USFP’s two-stage insertion design, which is intended to improve signal integrity and reduce contact wear. The interface is designed to support power dissipation of up to 100 watts without the need for a dedicated cooling circuit in the transceiver.
After developing the USFP technology, Nvidia and Cisco began working with TE Connectivity and Amphenol to create the remaining components of the new network infrastructure. Third-party companies are manufacturing cages, connectors and modules for the USFP interface, and other partners are expected to join the project.
New data center projects are deploying various networking solutions to handle their massive AI workloads. Cisco acknowledged that pluggable modules like USFP will be part of a broader ecosystem that includes near-packaged, co-packaged and other new optical architectures. These silicon photonics technologies could continue to bring significant benefits to the data center market even when the AI bubble eventually bursts.
