Products & Solutions

800G-Optimized Ethernet and VoQ Fabrics for AI/ML Workloads

Power-efficient and scalable designs.

The Hype of Generative AI

AI Has Drastically Accelerated the Size and Network Bandwidth Required in AI/ML Clusters With the Number of Compute Nodes and Accelerators Growing Significantly.

For privacy reasons YouTube needs your permission to be loaded. For more details, please see our Privacy Policy.
AI Networks
Demand High-capacity, Flat Systems for Efficient Handling of Extensive Data With Low Latency and High Throughput.
AI/ML workload
Organizations Adopt Ethernet-based or VoQ-based Fabrics, Aiming to Reduce Job Completion Times.
AI infrastructure efficiency
With Ethernet Offering Reliability and VoQ reducing Processing Delays.

Enable AI/ML Workloads

Power-efficient and scalable designs

800G AI/ML Optimized

Broadcom StrataXGS® Tomahawk® 5 Series

Ethernet Fabric

Accton’s 51.2 Tbps Broadcom StrataXGS® Tomahawk® 5 Series-based System

What’s Included
  • 51.2Tbps

  • 64x800G or 128x400G

  • High-radix

  • Robust and Compact

800G AI/ML Optimized

Broadcom StrataDNX™ Jericho3-AI & Ramon3

VoQ Fabric

Accton’s Broadcom Jericho3-based Network Cloud Packet Forwarder (NCP)

  • 14.4Tbps Full-duplex

  • 18x800G OSFP Network Interface Ports

  • 20x800G OSFP Fabric Interface Ports

Accton’s Broadcom Ramon3-based Network Cloud Fabric (NCF) Engine

  • 51.2 Tbps Switching

  • 21.6B Cell Switches Per Second

  • Cell-based Switching

Accton’s Tomahawk® 5 Series Ethernet Fabric

Robust, Compact, AI/ML Optimized

51.2T 800G – 2RU Variant

51.2T 800G 2RU Side

51.2T 800G 2RU Rear

51.2T 800G 2RU Front

51.2T 800G – 4RU Variant

51.2T 800G 4RU Side

51.2T 800G 4RU Rear

51.2T 800G 4RU Front

51.2 Tbps

Bandwidth In a Single Device

64x800G or 128x400G

Configuration Through Transceiver Type

2X

Throughput Increased

5nm

Process Technology

Not sure where to start?

We’ll help you figure it out 🚀

For privacy reasons YouTube needs your permission to be loaded. For more details, please see our Privacy Policy.

Overview

51.2Tbps Ethernet switching bandwidth. 2X increase in the switching throughput of prior generation switch platforms deployed

Single 51.2Tbps silicon design with OSFP and QSFP-DD interface variants

  • Two variants to support OSFP and QSFP-DD interface to optimize customer’s transceiver investment and solution preference

Support 64x 800G or 128x 400G Configuration through Transceiver Type

  • Support 64x 800G or 128x 400G configurations with the same hardware for superior scale out to 128x 400G through 2x 400G transceivers such as 2xFR4 and 2xDR4

High System Availability, Reliability, and superior Manufacturability

  • Power and fan tray redundant and advanced system design without using flyover cables to ensure achieve high availability, high reliability, and superior manufacturability

Ethernet Open Ecosystem

  • Open networking for disaggregated architecture for diverse deployment scenarios (SONiC and 3rd party software)
  • High performance Ethernet fabric for compute-intensive AI/ML workloads

Single 51.2Tbps silicon design with QSFP-DD interface

  • High-radix 128x 400G design to scale out Ethernet fabric through 400G and 4x 100G breakout cables or transceivers

Support large fanout through LR/FR/DR and 4x 100G QSFP-DD112 breakout

  • Support 128x 400G configurations for superior scale out through Transceivers such as LR and 4x 100G breakout solutions.

High System Availability, Reliability, and superior Manufacturability

  • Power and fan tray redundant and advanced system design without using flyover cables to ensure achieve high availability, high reliability, and superior manufacturability

Ethernet Open Ecosystem

  • Open networking for disaggregated architecture for diverse deployment scenarios (SONiC and 3rd party software)
  • High performance Ethernet fabric for compute-intensive AI/ML workloads

Not sure where to start?

We’ll help you figure it out 🚀

Request Demo

Request Consultation

Resources

AS9817-32D

800G Data Center Switch 1RU Variant

AS9817-64D

800G Data Center Switch 2RU Variant

DCS8500-128D

128x400G Data Center Switch 4RU Variant

AS9817-32O

800G Data Center Switch 1RU Variant

AS9817-64O

800G Data Center Switch 2RU Variant

Accton’s Jericho3 and Ramon3 VoQ Fabric Systems

Efficient and High-availability DDC Cluster

14.4Tbps Full-duplex Jericho3-AI Platform

VoQ-based Cell Fabric DDC Network Cloud Cluster

  • VoQ-based cell fabric DDC Network Cloud Cluster optimized to support AI/ML applications

High Capacity with Easy Migration from 400G to 800G GPU Clusters

  • 14.4 Tbps full-duplex switching capacity and features 18x 800G OSFP network interface ports and 20x 800G OSFP fabric interface ports to support 400G GPU clusters today and software upgrade to 800G GPU clusters

High System Availability, Reliability, and superior Manufacturability

  • Power and fan tray redundant and advanced system design without using flyover cables to ensure achieve high availability, high reliability, and superior manufacturability

Ethernet Open Ecosystem

  • Power and fan tray redundant and advanced system design without using flyover cables to ensure achieve high availability, high reliability, and superior manufacturability

Jericho3-AI Front

Jericho3-AI Rear

Jericho3-AI Front Panel

VoQ-based

Cell Fabric DDC Network Cloud Cluster

High Capacity

Easy Migration from 400G to 800G GPU Clusters

High System

Availability, Reliability, and superior Manufacturability

Ethernet

Open Ecosystem

Dual Ramon3 128x800G Compact Fabric Engine

High-radix 128 x 800G design with two 51.2Tbps self-routing cell-based silicon

  • High-radix 128x 800G design with two 51.2Tbps self-routing cell-based silicon for superior scale out DDC architecture

Chassis Design/Pay-as-you-grow

  • Fabric couples with the Jericho3-AI platforms in small, medium, and large clusters enabling a pay-as-you-grow model to bring down the TCO

High System Availability, Reliability, and superior Manufacturability

  • Power and fan tray redundant and advanced system design without using flyover cables to ensure achieve high availability, high reliability, and superior manufacturability

Ethernet Open Ecosystem

  • Open networking friendly for disaggregated architecture for compute-intensive workloads with 3rd party software

 

Dual Ramon3 Fabric Front Angle

Dual Ramon3 Fabric Rear

Dual Ramon3 Fabric Front Panel

51.2 Tbps

High-radix 128 x 800G design

Pay-as-you-grow

enabling cost scalability as your needs expand

High System

Availability, Reliability, and superior Manufacturability

Ethernet

Open Ecosystem

Multi-Stage Distributed Disaggregated Chassis (DDC)

Advantages

  • Perfect load balancing – No congestion
  • Single routing domain – Flat network
  • Scale-out as you grow – High radix

Logical Partition

  • Jericho3 (NCP: Network Cloud Packet Forwarder)
    • Switching, routing, queuing, QoS
    • DDC-level scheduling
  • Ramon3 (NCF: Network Cloud Fabric Engine)
    • Efficient cell-based switching
    • Low power, low latency

NCPs can be placed at

  • Compute /GPU racks, as ToR
  • Network rack, with NCFs

Further Scale with Multi-stage DDC Cluster

  • Single-stage cluster is limited up to 256 Jericho3 devices
  • The multi-stage cluster can further extend to 2K Jericho3 devices
  • Max 32Kx 800G ports in a single routing domain

Not sure where to start?

We’ll help you figure it out 🚀

Resources

ASA926-18XKE

14.4Tbps Full-Duplex Jericho3-AI Platform

AS9936-128D

Dual Ramon3 128x800G Compact Fabric Engine