NVIDIA 开放 AI 工厂设计，携手 Cisco 推出 USFP 高速网络模块
NVIDIA Opens AI Factory Designs, Teams Up With Cisco on High-Speed Networking
NVIDIA 与 Cisco 在 OCP 全球峰会上发布 USFP 光模块标准，每通道支持最高 400 Gb/s 电信号，带宽密度达现有 200 Gb/s 每通道光模块的 4 倍，能效约为其 1.3 倍。首批模块采用 16 个全双工通道，总容量最高 12.8 Tb/s，并计划扩展至 32 通道以上。NVIDIA 同时通过 MGX 架构开放 Vera Rubin 机架与计算托盘设计，Meta 将采用其 Spectrum-6 以太网交换机。
正文
当前显示原文，中文全文译文尚未就绪。
NVIDIA and Cisco have announced a new networking standard for AI data centres, designed to connect more computing systems without expanding existing network infrastructure.
The companies unveiled the Ultra-Dense Small Form Factor Pluggable (USFP) module at the Open Compute Project (OCP) Global Summit in Santa Clara, California. USFP supports electrical signals of up to 400 gigabits per second (Gb/s) per lane.
According to the companies, the technology offers four times the bandwidth density and around 1.3 times the power efficiency of existing 200 Gb/s-per-lane optical modules. NVIDIA and Cisco are developing the standard with TE Connectivity and Amphenol to support wider industry adoption.
The announcement forms part of NVIDIA’s wider push to expand open standards for AI infrastructure. At the summit, the company also contributed designs for its Vera Rubin rack and compute trays through its open MGX architecture.
As AI models grow and demand for AI inference increases, data centres must connect more computing systems while managing power consumption, costs and network complexity. Traditional expansion can require additional switches and optical links, increasing latency and energy use.
USFP aims to address these challenges by increasing the amount of data that networks can carry within the same physical footprint. The first planned modules will feature 16 full-duplex lanes in a design similar in size to the widely used Octal Small Form Factor Pluggable (OSFP) module.
The companies plan to expand the design to 32 lanes or more.
The module can deliver up to 12.8 terabits per second (Tb/s) in total capacity. This could help AI infrastructure operators connect more computing resources without adding network layers or overhauling existing data centre infrastructure.
“Building the world’s largest AI factories starts with rethinking the smallest of connections,” said Gilad Shainer, Senior VP of Networking at NVIDIA.
He said companies were building an ecosystem to connect more computing systems more efficiently and support the next generation of AI factories.
NVIDIA also announced software components for managing AI infrastructure and worked with industry partners on standards for data centre design and power grid resilience.
Meta is adopting NVIDIA’s Spectrum-6 Ethernet switches for its AI infrastructure, while NVIDIA and Google announced support for Google’s Falcon transport protocol on future NVIDIA networking products.
As for USFP, Cisco said the technology would allow AI infrastructure to scale without adding operational complexity. “It is designed to deliver a new level of bandwidth density while preserving the serviceability, deployment flexibility and operational simplicity of pluggable optics,” said Bill Gartner, Senior VP and GM of Optical Systems and Optics at Cisco.
The design also supports up to 100 watts of power dissipation without requiring a separate, dedicated cooling circuit within the transceiver. Its connector design aims to reduce contact wear and improve signal reliability.
NVIDIA and Cisco are working with TE Connectivity and Amphenol to develop the module’s cages, connectors and optical components. The companies plan to advance USFP through a community-led, multi-source project, allowing suppliers to develop compatible hardware.