Specifications
Ethernet Fabric Silicon
Broadcom StrataXGS Tomahawk 5, 51.2 Tbps per device, 5nm process
Ethernet Port Config
64x800G or 128x400G via transceiver type on the same hardware
1RU 800G System
AS9817-32O and AS9817-32D: 800G data center switches, 1RU variant
2RU 800G System
AS9817-64O and AS9817-64D: 64x800G OSFP or QSFP-DD, single 51.2T silicon
4RU 400G System
DCS8500-128D: high-radix 128x400G QSFP112, compatible with QSFP56 and QSFP28
VoQ NCP Platform
ASA926-18XKE: Broadcom Jericho3-AI Network Cloud Packet Forwarder
NCP Capacity
14.4 Tbps full-duplex switching capacity
NCP Ports
18x800G OSFP network interface ports plus 20x800G OSFP fabric interface ports
NCF Engine
Broadcom Ramon3 Network Cloud Fabric engine, 51.2 Tbps switching capacity
Dual-Ramon3 Fabric
6RU compact fabric engine, 128x800G (dual 51.2 Tbps) fabric port switch
Cell Switching Rate
21.6 billion self-routing cell switches per second
Cluster Scale
Two-stage DDC network architecture builds a 32K-GPU 800G/400G AI/ML cluster
Migration Path
Deploy 400G GPU clusters now and migrate to 800G via software upgrade without replacing switches
DDC Architecture
Perfect load balancing, single routing domain, flat network, scale-out by high radix
Scale Limits
Single-stage up to 256 Jericho3 devices; multi-stage extendable to 2K devices; 32K x 800G ports in one domain
Software
Open networking with SONiC and third-party disaggregated software
Redundancy
Redundant power and fan trays; advanced system design without flyover cables for five-nines availability
Availability
Tomahawk 5 fabric first customer shipment July 2023; Jericho3/Ramon3 VoQ fabric GA Q4 2023; current lineup shipping in 2026
Overview
Accton Technology's 800G AI/ML fabrics give hyperscale and enterprise operators both an Ethernet-based fabric and a Virtual Output Queue (VoQ) based Distributed Disaggregated Chassis (DDC) fabric, engineered to cut job completion time for large language model training and inference clusters.
The Ethernet fabric pairs a single 51.2 Tbps Broadcom Tomahawk 5 device with 64x800G or 128x400G ports across compact 1RU, 2RU and 4RU systems. Power and fan trays are redundant and the systems are built without flyover cables, preserving signal integrity and supporting five-nines availability.
For the largest clusters, the VoQ DDC fabric couples Jericho3-AI packet forwarders (NCP) with Ramon3 cell-based fabric engines (NCF). The two-stage design scales to 32K x 800G ports in a single routing domain with perfect load balancing, and lets operators start on 400G and move to 800G with a software upgrade only.
Key Benefits
High radix, low latency, and investment protection. A single flat routing domain, cell-based load balancing, and a 400G-to-800G software migration path reduce total cost of ownership for AI/ML clusters.
Applications
Hyperscale AI/ML training and inference clusters, GPU cluster scale-out fabrics, distributed disaggregated chassis (DDC) networks, and next-generation cloud data center networking.
Request a Quote — ACCTON 800G AI/ML FABRICS — TOMAHAWK 5 AND JERICHO3/RAMON3 DDC SWITCHING
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →