@Scale: Networking 2026
| title | author | length |
|---|---|---|
| @Scale Networking Event Welcome | Vignesh Vijayanath | 04m |
Welcome and orientation for At Scale Networking 2026 describe a conference that centers on building and operating global data center networks and the physical… | ||
| Keynote | Omar Baldonado | 23m |
Meta reports concrete scale: 1.2 million H100s (~zetaflop class), a 129k‑GPU cluster and repeated 24k training clusters, plus rapid‑deployment Prometheus… | ||
| Resilient AI Supercomputer Networking using MRC and SRv6 | Jitendra Padhye | 22m |
A new RDMA transport called MRC sprays packets across many paths and uses SRv6 source routing so hosts bypass failures without switch control, handling… | ||
| MetaRoCE: Meta’s RDMA Transport | Arvind Srinivasan, Kingshuk Mandal | 20m |
MetaRoCE is a clean‑sheet RDMA transport that sprays packets across many paths, keeps per‑path RTT/congestion state and per‑path retransmission timers, and… | ||
| MetaRoCE: Meta’s RDMA Transport | Arvind Srinivasan, Kingshuk Mandal | 22m |
MetaRoCE is a clean‑sheet RDMA transport that Meta builds for clusters scaling to over a million GPUs, claiming 90% efficiency via per‑packet multipath… | ||
| Anatomy of @Scale Machine Learning Network Operations | Jim Julson | 20m |
Claims that AI/ML training must run at line rate across 128,000 GPUs with zero flaps, so CoreWeave builds blueprints around four cornerstones: people… | ||
| Anatomy of @scale Machine Learning Network Operations | Jim Julson | 21m |
Operations at Coreweave target synchronously connecting 128,000 GPUs with zero packet loss across millions of ports and tens of thousands of miles of fiber, so… | ||
| MetaRoCE: From Spec to NIC to Open Source | Balakrishnan Raman, Sandeep Nagaraj | 16m |
MetaRoCE implements a receiver-driven, multipath RDMA stack with a software reference, compliance suite, and an AMD P4 programmable NIC implementation… | ||
| MetaRoCE: From Spec to NIC to Open Source | Balakrishnan Raman, Sandeep Nagaraj | 19m |
MetaRoCE ships a full software reference implementation and compliance suites and runs on an AMD P4 programmable AI NIC to prototype features before silicon… | ||
| Live Group Q&A Session | James Zeng, Arvind Srinivasan, Kingshuk Mandal, Sandeep Nagaraja, Balakrishnan Raman | 18m |
Meta builds a new transport called metari because RoCEv2 cannot tolerate out‑of‑order multipath delivery; a working reference implementation with AMD and… | ||
| Live Panel: Scale-Out Networking for AI Supercomputers | Joseph Provine, Andrew Dickinson, Jim Julson, Jitendra Padhye | 31m |
Panel describes how real supercomputers tie tens of thousands of accelerators into one fabric and reports implementing MRC across vendors to get synchronized… | ||
| AI for NI | Mohab Gawish | 09m |
Network capacity grows from 1.3 GW to 7 GW (≈5x) while heterogeneity across GPU generations, NICs, cooling and power makes complexity grow exponentially, not… | ||
| Lightning Talks Introduction - AI for NI | Mohab Gawish | 10m |
Describes a pipeline that turns raw AI demand into an operational network and claims AI will bridge the divergence gap between exponentially growing complexity… | ||
| The Plan That Learns | Ben Vallis | 08m |
AI agents triage hundreds of emerging traffic patterns down to a handful that need human time, enabling action inside narrow buy windows (weeks down to… | ||
| The Plan That Learns | Ben Vallis | 10m |
Network capacity planning uses AI to triage hundreds of weekly novel traffic patterns into a handful that need action, shrinking investigation windows from… | ||
| 10x Network Design Engineer | Maaz Mohiuddin | 10m |
Datacenter designs become codified rules, retrieval of hundreds of past designs, and a fleet of specialist agents that size fabrics, render configs, and… | ||
| 10x Network Design Engineer | Maaz Mohiuddin | 10m |
Describes a system that turns capacity briefs into build‑ready datacenter designs using multiple LLM agents plus deterministic rule engines and a validation… | ||
| AI-Native Network Capacity Management | Mohab Gawish | 10m |
Claims an LLD→RPL translation agent that reads low-level design documents, extracts routing intent, and authors RPL so the existing RPL + deterministic… | ||
| AI-Native Network Capacity Management | Mohab Gawish | 11m |
AI agents replace the manual step that turns messy design artifacts into device-ready config by parsing text, diagrams and tables and emitting a structured RPL… | ||
| AI-Native Network Operation | YuLing Chen | 11m |
Meta builds a hierarchical, parent-orchestrated agent system that runs in production to detect, root-cause and mitigate incidents, claiming reliability gains… | ||
| AI-Native Network Operation | YuLing Chen | 12m |
Meta describes deploying hundreds of thousands of agents in a hierarchical system that already detects incidents standard monitoring misses, traces root causes… | ||
| Physical Network Design at Scale | Brandon Premo, Richard Cziva | 20m |
Loom automates end-to-end physical network design: Planogram places thousands of devices onto racks and failure domains using lexicographic optimization (four… | ||
| Physical Network Design at Scale | Brandon Premo, Richard Cziva | 20m |
Meta describes Loom, an end-to-end system that automates physical network design: Planogram places every network device to a precise rack and row using an ILP… | ||
| Building Scalable AI Cloud Interconnects with SRv6 | Alexey Gorovoy | 18m |
Nebius uses SRv6 to replace VXLAN/MPLS stitching and run a single IPv6 routed fabric that delivers end-to-end VPN and Traffic Engineering without protocol… | ||
| Building scalable AI cloud interconnects with SRv6 | Alexey Gorovoy | 18m |
Nebius replaces stitched VXLAN+MPLS stacks with an SRv6-first design that provides end-to-end Layer‑3 VPNs and traffic engineering across data center and WAN… | ||
| Fiber Networks Innovations for AI at Scale | Fabrice Ouandji, Sebastian Gault | 19m |
Project delivers over 100,000 strands per path by jetting hundreds of cables instead of buying paths from carriers, using 148 864‑fiber cables in phase one and… | ||
| Fiber Networks Innovations for AI at Scale | Fabrice Ouandji, Sebastian Gault | 20m |
Meta needs to escalate from ~864 to 100,000 fiber strands between nearby data centers, so teams design and dig mega duct banks with 20 ft deep, 9 ft wide… | ||
| Virgo: Scale-out Data Center Network for AI | Jeongkeun "JK" Lee | 20m |
Virgo is a dedicated accelerator-to-accelerator fabric: a flat two-layer non-blocking topology built from high-radix switches and skinny links (e.g., 200G 30s… | ||
| Virgo: Scale-out Data Center Network for AI | Jeongkeun "JK" Lee | 19m |
Virgo is a dedicated accelerator-to-accelerator fabric that collapses to a flat two-layer non‑blocking topology using high‑radix switches and 200Gbps 30C… | ||
| Co-Designing Communication for AI Accelerators | Rajeev Nair, Wes Bland | 18m |
MTIA 300 integrates message engines and near-memory compute to offload communication from compute cores, running RDMA via Broadcom NIC banks (up to 1.2 TB/s… | ||
| Co-Designing Communication for AI Accelerators | Rajeev Nair, Wes Bland | 18m |
Meta describes a custom accelerator MTIA 300 with on-chip message engines and near memory compute that offload communications into HBM and avoid PE and host… | ||
| Live Panel: Building at the Roofline — Planning and Delivering Gigawatt-Scale AI Infrastructure | Omar Baldonado, Amrita Kenue, Lucas Benson, Nikita Parikh | 27m |
Plans target gigawatt-scale fleets with company capex comparable to a national GDP and a planning model split into two planning cycles (0–18 months and 3–5… | ||
| The Future of AI Networking Starts Here | 01m | |
Demonstrates a full‑stack debugging workflow that links on-host traces with switch sFlow and hardware counters to pinpoint interconnect congestion, identifies… | ||
| nothing matches | ||