Power-efficient and scalable designs with a 51.2 Tbps Ethernet fabric of 64x800G ports, a 51.2 Tbps Ethernet fabric of 128x400G ports, and a VoQ-based fabric Distributed Disaggregated Chassis (DDC) architecture with a 128x800G Network Cloud Fabric (NCF) engine providing the best alternative options for AI/ML workloads.
Accton Technology Corporation (Accton) (TWSE: 2345), a leading provider of networking and communications solutions, today announced a series of 800G-optimized products that can provide Ethernet-based fabric and Virtual Output Queue (VoQ) based fabric for AI/ML workloads. The announcement includes a 2RU 51.2 Tbps Ethernet-based fabric system with 64x800G ports, a 4RU 51.2 Tbps Ethernet-based fabric system with 128x400G ports, and a VoQ-based fabric Distributed Disaggregated Chassis (DDC) architecture with a 128x800G Network Cloud Fabric (NCF) engine.
The hype of generative AI has drastically accelerated the size and network bandwidth required in AI/ML clusters with the number of compute nodes and accelerators growing significantly. AI networks require high-capacity systems in a flat architecture that can handle large amounts of data with low latency and high throughput. To meet the demands of AI/ML workloads, either Ethernet-based fabrics or VoQ-based fabrics are being adopted to reduce job completion time.
Accton’s 51.2 Tbps Broadcom StrataXGS® Tomahawk® 5 series-based system with 64x800G or 128x400G ports provides a high-radix, deployment-friendly Ethernet-fabric. The 2RU system design is robust and packed in a compact form factor with power and fan tray redundancy to achieve five-nines high availability and a wide environmental operating range for data center cloud applications. With a design that does not use flyover cables, the system offers known good system quality and reliability while preserving the port assignment flexibility for customers. The platform comes in two variants, OSFP800 or QSFP-DD800 interface options, for flexible deployment supporting passive copper DAC on all ports and long-distance ZR+ optics. The 4RU 51.2 Tbps Tomahawk 5 series-based 128x400G system provides high-radix 200G/400G QSFP112 (compatible with QSFP56 and QSFP28) connectivity to accelerators and compute nodes in a flat architecture that reduces latency and required power, which enables networks to be scaled-out sustainably.
Accton’s Broadcom StrataDNX™ Jericho3-AI and StrataDNX™ Ramon3 systems are designed to form a VoQ-based fabric DDC Network Cloud Cluster that is optimized to support AI/ML applications. The Broadcom Jericho3-AI based Network Cloud Packet Forwarder (NCP) supports a 14.4 Tbps full-duplex switching and features 18x800G OSFP network interface ports and 20x800G OSFP fabric interface ports. The Broadcom Ramon3-based Network Cloud Fabric (NCF) engine supports a 51.2 Tbps switching capacity. The cell-based switching eliminates the Ethernet overhead and can effectively load balance all fabric links to build an efficient and high-availability DDC cluster.
With Accton’s 6RU, dual-Ramon3, 128x800G (dual-51.2 Tbps) fabric port switch and 2RU, Jericho3-AI, 18x800G (14.4 Tbps leaf) Ethernet network interface port switch, customers can build a 32K-GPU 800G/400G AI/ML cluster with a two-stage DDC network architecture to enable 400G GPU clusters now and migrate to 800G GPU clusters later with a software upgrade without replacing the switches. This provides excellent cost saving and investment protection that reduces capital spending while providing the flexibility for customers to “pay as you grow” without large upfront cost to enable AI and ML applications.
Accton takes the hardware design further for the dual-Ramon3 system by optimizing the port mapping within the box, achieving traffic load balancing that accommodates extreme traffic conditions. The innovative design approach reduces the restrictions for network architects to balance traffic loads. With more than thirty years of networking equipment design experience, Accton’s Ramon3 and Jericho3-AI systems offer much lower system power consumption, reduced latency, minimized port-to-port timing skews, and high manufacturability with no flyover cables.
Michael K.T. Lee, Sr. Vice President of Research and Development Center, at Accton said, “Broadcom’s Tomahawk 5 and Roman3/Jericho3-AI form the building blocks of Accton’s 800G-optimized products for next-generation AI/ML clusters. We are proud of Accton’s engineers who have excelled in the innovative design and development of these groundbreaking 800G systems. The significant increase in radix and throughput offered by the 800G AI fabric together with the reduced power consumption delivers a performance that will push the boundaries of efficiency in AI clusters.”
Wei-Ai Tai, director, business development , Core Switching Group, Broadcom said, “With the availability of Tomahawk 5 and Ramon3/Jericho3-AI based fabrics, hyperscalers will want to upgrade their AI infrastructure to cope with the surge in bandwidth demand as well as improve AI cluster performance and enable a lower total cost of ownership.”
Accton’s Tomahawk-based Ethernet fabric has already achieved First Customer Shipment (FCS) in July. The Jericho3/Ramon3 VoQ-based fabric will achieve General Availability (GA) in Q4 2023.
Visit AITechPark for cutting-edge Tech Trends around AI, ML, Cybersecurity, along with AITech News, and timely updates from industry professionals!