Follow on Twitter using the following hashtags or usernames: #NFD40
Watch on YouTube
Watch on Vimeo
As AI training and inference scale, the network must function as an extension of the compute fabric. This session explores the architectural requirements for high-performance AI data centers. We will examine the shift toward deterministic networking to mitigate tail latency and fabric congestion, alongside critical hardware innovations — including advanced cooling and next-generation optics, designed to maximize performance and power efficiency. Attendees will gain technical insights into building a unified, programmable fabric that optimizes performance and scalability for high-density AI environments.
The presentation emphasizes that an AI-ready data center requires simultaneous innovation across five key dimensions: scalability, power efficiency, security, operational management, and silicon diversity. Cisco highlights the rapid transition in networking speeds, moving from 400G and 800G to 1.6T in just two years to keep pace with GPU evolution. A major focus is placed on the shift toward Ethernet for scale-out fabrics, as it offers a consistent operational model across front-end, back-end, and management networks. To achieve performance parity with InfiniBand, Cisco utilizes its Silicon One architecture, featuring deep, fully shared packet buffers and programmable pipelines that allow for the mid-cycle introduction of advanced features like dynamic load balancing and packet spraying to mitigate microbursts and reduce job completion time.
Cisco also detailed its strategic partnership with NVIDIA, which goes beyond simple reselling to include co-engineering systems that integrate Cisco’s NXOS and Nexus Dashboard with NVIDIA’s Spectrum-4 silicon. This collaboration aims to provide repeatable, standardized reference architectures that support high-performance features like adaptive routing and direct data placement. Furthermore, the discussion introduced the concept of “scaling across” geographically distant data centers, necessitating P-series silicon with deeper buffers and advanced optics for long-haul connectivity. By offering a vertically integrated stack encompassing silicon, hardware, operating systems, and optics, Cisco aims to provide a cohesive and programmable fabric that addresses the extreme power and performance demands of modern agentic AI workloads.
Personnel: Faraz Taifehesmatian, Richard Licon
Watch on YouTube
Watch on Vimeo
As AI training and inference scale, the network must function as an extension of the compute fabric. This session explores the architectural requirements for high-performance AI data centers. We will examine the shift toward deterministic networking to mitigate tail latency and fabric congestion, alongside critical hardware innovations — including advanced cooling and next-generation optics, designed to maximize performance and power efficiency. Attendees will gain technical insights into building a unified, programmable fabric that optimizes performance and scalability for high-density AI environments.
The presentation details Cisco’s strategic use of its Silicon One architecture, specifically the G-Series for AI scale-out and the P-Series for “scale-across” data center interconnects. The G-Series, highlighted by the G200 and G300 ASICs, provides high-radix connectivity with up to 512 ports of 200G and fully shared packet buffers to eliminate the performance bottlenecks found in traditional slice-based architectures. A core focus is the Cisco Intelligence Packet Flow (IPF), which enables advanced load balancing techniques such as packet spraying and flowlet switching. These features allow Ethernet to mimic the lossless properties of InfiniBand, ensuring high job completion times for RDMA-heavy AI workloads while maintaining a programmable pipeline that can adapt to evolving standards like Ultra Ethernet mid-cycle.
Hardware innovation is further demonstrated through new form factors and cooling solutions designed for high-density AI environments. Cisco introduced liquid-cooled chassis, such as the N9364F-SG3-L, which achieves 100% liquid cooling to handle the massive power requirements of 100-terabit ASICs without the need for fans. These systems support next-generation optics, including Linear Pluggable Optics (LPO) that reduce power consumption by half and coherent ZR/ZR+ optics for long-haul connectivity up to 1,000 km. Additionally, Cisco’s partnership with NVIDIA was underscored through the N9100 series, which integrates NVIDIA Spectrum-4 and Spectrum-6 silicon into the Cisco ecosystem. This gives customers the choice between a vertically integrated Cisco fabric or an end-to-end NVIDIA Spectrum-X solution, all managed through a consistent operating system and the Nexus Dashboard.
Personnel: Faraz Taifehesmatian
Watch on YouTube
Watch on Vimeo
As AI training and inference scale, the network must function as an extension of the compute fabric. This session explores the architectural requirements for high-performance AI data centers. We will examine the shift toward deterministic networking to mitigate tail latency and fabric congestion, alongside critical hardware innovations — including advanced cooling and next-generation optics, designed to maximize performance and power efficiency. Attendees will gain technical insights into building a unified, programmable fabric that optimizes performance and scalability for high-density AI environments.
The presentation introduces the third generation of Cisco’s bidirectional (BiDi) technology, specifically the 400G BiDi optic. This innovation addresses fiber infrastructure constraints by enabling fiber reuse, allowing customers to upgrade from 40G or 100G to 400G over existing duplex multi-mode fiber without installing new trunk cables or patch panels. By utilizing four wavelengths at 100G each over a single fiber pair, the 400G BiDi simplifies the physical layer with LC connectors, making it eight times more fiber-efficient than parallel SR8 solutions. This approach offers significant financial and operational benefits for both brownfield and greenfield deployments by reducing installation costs and troubleshooting complexity.
A major portion of the session focuses on the critical role of optics reliability and Cisco’s advanced silicon photonics in AI environments. Unlike traditional networks where retransmissions are common, AI workloads are highly synchronized; a single unreliable optical link can cause GPU clusters to stall, potentially reducing performance by 40%. Cisco’s silicon photonics architecture integrates electronics and photonics into a single system, improving stability and power efficiency for 800G and 1.6T scales. Notable highlights include the 1.6T pluggable optic, which supports flexible breakout options, and the 800G Linear Pluggable Optic (LPO). By removing the DSP from the optic and shifting signal conditioning to the switch ASIC, the LPO solution reduces power consumption by 50% per module and lowers overall system latency, providing a more reliable and sustainable foundation for large-scale AI factories.
Personnel: Paymon Mogharabi
Thank you for being part of the Tech Field Day community! Our mailing list is a great way to stay up to date on our events and technical content, and we appreciate your signup.
We promise that we’ll never spam you, send ads, or sell your information. This list will only be used to communicate with our community about our events and content. And we’ll limit it to no more than one message per week.
Although we only need your email address, it would be nice if you provided a little more information to help us get to know you better!
