Digital Plumber Plumbing the information age
RSS

Topics

Data Center & AI Networking

The networks under AI: Ethernet and InfiniBand fabrics, Ultra Ethernet, GPU cluster networking and 800G/1.6T optics.

168stories in the archive
9 up from 5last 7 days · 5 the 7 before
27last 30 days
Jun 11first covered · 2026

Recent activity

Stories per week

Latest stories

10 most recent
  1. IEEE ComSoc Blog · Nov 16, 2026 · Analysis

    Broadcom pitches disaggregated Ethernet fabric against Cisco Silicon One

    Broadcom unveiled its open Ethernet fabric portfolio at SC26, contrasting its disaggregated ASIC strategy (Tomahawk Ultra, Tomahawk 6, Jericho 4 paired with Thor Ultra NICs) against Cisco's unified P4-programmable Silicon One architecture for AI workloads.

    Why it matters Clarifies architectural divergence in 102.4T switching: disaggregated ASICs versus unified silicon affects tail-latency behavior, fabric programmability, and multi-vendor interoperability for large AI clusters.

    Nov 20, 2026 edition

  2. Communications Today · Nov 18, 2026 · Industry news

    Arista launches R4 routers for edge AI inferencing

    Arista unveiled R4 routers—compact, hardened access routers, coherent transport router, and 25.6-Tbps pluggable optical line system—designed for distributed AI inferencing at hyperscalers and neocloud operators needing 1.6T links with liquid cooling.

    Why it matters Extends AI networking from data-center spine-leaf to metro/edge; enables neoclouds and hyperscalers to link 100k-1M accelerator clusters across multiple facilities with native 1.6T interconnect.

    Nov 20, 2026 edition

  3. SDxCentral · Nov 11, 2026 · Industry news

    Oracle's OCCL plugin lifts NCCL collective bandwidth up to 14%

    Oracle released Oracle Cloud Infrastructure Collective Communications Library (OCCL), a network plugin that improved NCCL AllReduce bandwidth by up to 10% and AllToAll by 14% on 16–128 node GB300 clusters, deployed atop OCI's multiplanar RDMA architecture with in-transport congestion control and NIC failure recovery.

    Why it matters OCCL provides production-tested collective communications offload for OCI GPU clusters at scale, reducing network-induced delays in multi-node all-reduce patterns; multiplanar redundancy and transport-layer congestion control narrow the performance gap between OCI Spectrum-X Ethernet and proprietary NVIDIA Quantum InfiniBand on NCCL workloads.

    Nov 18, 2026 edition

  4. HPCwire · Nov 11, 2026 · Industry news

    Broadcom pushes Scale-Up Ethernet Transport for multi-vendor AI fabrics

    Broadcom hosted SC26 workshop coverage of open AI and HPC fabric standards, emphasizing Scale-Up Ethernet Transport (SUE-T) under Open Compute Project and Ultra Ethernet Consortium specifications for deterministic, lossless transport, load balancing, congestion control, resilience, and telemetry across multi-vendor environments.

    Why it matters SUE-T standardization under Open Compute Project and UEC 1.0 certification enable practitioners to architect multi-vendor AI fabrics with deterministic transport guarantees and certified interoperability; vendors can now ship UEC-compliant fabric switches, NICs, and telemetry stacks without proprietary lock-in.

    Nov 18, 2026 edition

  5. TechBlog ComsOC · Nov 16, 2026 · Analysis

    Broadcom pitches disaggregated AI Ethernet fabric against Cisco Silicon One

    Broadcom presented its disaggregated AI fabric strategy at SC26 combining Tomahawk 6 (102.4 Tbps), Tomahawk Ultra (low-latency intra-rack), Jericho 4 (deep-buffer routing), and Thor Ultra 800G NIC with UEC compliance, contrasting with Cisco's unified Silicon One approach.

    Why it matters Network architects must choose between Broadcom's specialized role-based ASICs (optimized per function) versus Cisco's unified programmable silicon; Broadcom's approach reduces switching latency and power per function, critical for 100k+ GPU clusters.

    Nov 17, 2026 edition

  6. CASRAI · Nov 16, 2026 · Industry news

    Lambda runs co-packaged optics across 10,000 Blackwell GPUs

    Lambda and NVIDIA announced the first large-scale commercial deployment of co-packaged optics (CPO) networking, connecting over 10,000 Blackwell Ultra GPUs at Lambda's Quebec data center. The CPO fabric delivered 3.8% lower network latency at rack scale and 10.1% lower latency at pod scale compared to conventional pluggable optics.

    Why it matters CPO switches operating at 800 Gb/s achieve median optics temperature of 36°C versus 51°C for pluggables, reducing thermal density and enabling denser AI fabrics without data center infrastructure redesign.

    Nov 16, 2026 edition

  7. CASRAI · Nov 15, 2026 · Industry news

    Broadcom shows Tomahawk 6 and Thor Ultra at SC26

    Broadcom demonstrated Tomahawk 6 and Thor Ultra 800G Ethernet silicon at SC26, highlighting compatibility with Ultra Ethernet Consortium (UEC) specifications for AI workloads. Thor Ultra eRoCE NICs achieve job completion times matching 800G InfiniBand quantum fabrics by placing packets directly into GPU memory.

    Why it matters Open Ethernet silicon with UEC compliance and direct GPU memory placement eliminates packet ordering dependencies, allowing switches to spray packets across all links and closing the performance gap with InfiniBand at Ethernet scale and cost.

    Nov 16, 2026 edition

  8. Silicon Atlas · Nov 13, 2026 · Analysis

    Moving from 800G to 1.6T cuts module spending 25%

    AI clusters require dramatically more optical interconnect per unit of compute than previous data center generations, pulling transceiver speeds from 800G toward 1.6T and driving early adoption of co-packaged optics. Moving from 800G to 1.6T preserves total fiber capacity and halves the module count; at an assumed 50% price premium, module spending for the same capacity falls 25%.

    Why it matters Operators can reduce fiber run counts and total optics spend per data center while increasing port density, improving rack power budgets and reducing cooling load—the physics of AI density now favor 1.6T transceiver adoption.

    Nov 15, 2026 edition

  9. HPCwire · Nov 10, 2026 · Vendor release

    Oracle plugin lifts NCCL bus bandwidth up to 10%

    Oracle Cloud Infrastructure introduced Oracle Collective Communications Library (OCCL) network plugin for distributed AI workloads on OCI-supported GPUs and RDMA-enabled infrastructure. Benchmarks on NVIDIA GB300 systems show OCCL increased average NCCL bus bandwidth by up to 10% for AllReduce operations across 16-128 nodes.

    Why it matters OCCL addresses production constraints in large-scale AI training: congested network paths, failed links, and failed NICs interrupt collective communication and force training jobs to stall or restart—this plugin recovers automatically and reroutes around failures.

    Nov 15, 2026 edition

  10. Fierce Network / NextPlatform · Nov 5, 2026 · Industry news

    Arista splits AI fabrics between ESUN and UEC Ethernet

    Arista announced dual Ethernet standards for scale-up and scale-out: ESUN for scale-up (supporting up to 144 accelerators per rack, 1,024 cross-rack) and UEC-based Ethernet for scale-out. Arista has surpassed 100 Etherlink customers and forecasts providing networking for 10 gigawatts of AI cluster capacity by end of 2027, linking 10+ accelerator types.

    Why it matters Scale-up networking is now treated as a first-class citizen with full-featured network stack parity, enabling tighter GPU/XPU clustering and heterogeneous acceleration without proprietary lock-in—crucial for operators building multi-vendor AI fabrics.

    Nov 12, 2026 edition

Coverage history

158 earlier stories

November 2026 9

October 2026 41

September 2026 21

August 2026 35

July 2026 24

June 2026 28