NVIDIA Spectrum-X Network Platform Architecture
Presented by Gilad Shainer (Nvidia) | David Iles (Nvidia) The NVIDIA Spectrum-X Networking Platform is the first Ethernet platform designed specifically to improve the performance of Ethernet-based AI clouds. This technology achieves 1.7X improved AI performance for massive AI workloads such as LLM, along with 1.7X better power efficiency and predictable performance in multi-tenant environments. Spectrum-X is built on the tight coupling of the Spectrum-4 Ethernet switch with the NVIDIA BlueField-3 data processing unit (DPU). Spectrum-X platform is based on the combination of lossless Ethernet connectivity with fine grain adaptive routing and direct data placement for achieving nearly perfect effective bandwidth at load and scale. Telemetry based congestion control and special network delivery data path create the conditions of performance isolations for multi-tenant AI at scale. The session will provide a technical review of the AI Ethernet network innovations presented with Spectrum-X platform.

Source Routing for AI Fabrics

Keynote: Networking for AI and HPC, and Ultra Ethernet

Networking for AI Scaling, presented by Broadcom

NVIDIA Spectrum X vs. Arista & Broadcom: Is the AI Networking Boom Big Enough for Everyone?

NVIDIA’s New Photonic Technology Explained

NVIDIA Spectrum-X Network Platform Architecture

Accelerating multimodal LLMs development with NVIDIA’s Blackwell architecture

Intro to NVIDIA NIM for AI Builders

Networking for AI

NCCL: High-Speed Inter-GPU Communication for Large-Scale Training - Sylvain Jeaugey, NVIDIA

Building a GPU cluster for AI

NVIDIA Networking for HPC, AI, and Accelerated IO

Inside the World's Largest AI Supercluster xAI Colossus

Inside a NEW AI Cluster - Tour with NVIDIA B200

27 Aug 18: Webinar: Introduction to InfiniBand Networks

Fostering Collaboration Designing Data Centers for Tomorrows AI Workloads - presented by NVIDIA

How do Graphics Cards Work? Exploring GPU Architecture

Driving Performance: Efficient Networking for AI Infrastructure

NVIDIA didn't want me to do this

