SEMI VISION

SEMI VISION

Share this post

SEMI VISION
SEMI VISION
Broadcom’s CPO Strategy and Its Implications for the Future of Optical Interconnects

Broadcom’s CPO Strategy and Its Implications for the Future of Optical Interconnects

Original Articles by SemiVision Research (Broadcom)

SEMI VISION's avatar
SEMI VISION
Jun 04, 2025
∙ Paid
18

Share this post

SEMI VISION
SEMI VISION
Broadcom’s CPO Strategy and Its Implications for the Future of Optical Interconnects
5
Share

Broadcom’s CPO Approach & Industry Reactions

As artificial intelligence (AI) and high-performance computing (HPC) accelerate at an unprecedented pace, networking infrastructure is facing massive new challenges. Every generation of AI models demands exponentially more bandwidth, lower latency, and greater energy efficiency. In this context, Broadcom introduced the world’s first 102.4 Tbps Ethernet switch chip—Tomahawk 6 (TH6)—not only pushing the performance limits of Ethernet but also redefining the network fabric of the modern data center.

Built on TSMC’s advanced 3nm process, Tomahawk 6 is engineered to support ultra-large-scale AI compute clusters, satisfying both scale-up and scale-out topologies. Whether connecting hundreds of thousands of GPUs or enabling highly energy-efficient, ultra-low-latency switching fabrics, TH6 delivers a new level of flexibility and performance.

This isn’t merely a faster switch—it marks a strategic pivot in the evolution of Ethernet for the AI era. With growing adoption of open standards and emerging technologies such as Co-Packaged Optics (CPO), Broadcom is reshaping Ethernet’s role in hyperscale AI infrastructure. Tomahawk 6 signals that the next generation of AI data centers is ready.

Launched on June 3, 2025, Tomahawk 6 became the first single-die Ethernet switch chip to exceed 100 Tbps, solidifying Ethernet’s relevance as the backbone of AI data center networks. With double the switching capacity of its predecessor (Tomahawk 5), TH6 is now the highest-performance Ethernet switch available. The use of a 3nm process not only boosts transistor density but also significantly lowers power consumption—critical for large-scale, power-hungry AI deployments.

Designed to support over 1 million connected processors, TH6 is ideal for LLM training platforms, generative AI pipelines, and next-gen deep learning systems. Its dual support for scale-up and scale-out architectures enables flexible multi-tier switching fabrics, essential for the Zettascale AI era’s extreme demands on data exchange.

Unlike proprietary technologies like NVIDIA’s InfiniBand, Tomahawk 6 adheres to open Ethernet standards, supporting protocols like Ethernet/IP/UDP and aligning with widely adopted open software stacks such as SONiC. This openness gives hyperscalers and system builders greater flexibility, avoids vendor lock-in, and reduces total cost of ownership.

Technically, Tomahawk 6 is Broadcom’s first Ethernet switch to adopt a chiplet architecture—modularizing different functional blocks within the same package to increase manufacturing flexibility and yield. This architecture also paves the way for the adoption of Co-Packaged Optics (CPO).

In its CPO configuration, TH6 can be directly integrated with optical transceivers, minimizing electrical-to-optical signal distance, reducing I/O power consumption, and enabling future upgrades to 800G, 1.6T, and beyond. This addresses one of the critical bottlenecks in traditional pluggable optics: electrical signal loss and power inefficiency at high speeds.

Additionally, TH6 features Cognitive Routing 2.0 and Congestion-Aware Flow Control, allowing it to dynamically adjust data paths and intelligently route around network hotspots. These capabilities are essential for large-scale AI workloads, enabling the system to avoid incast congestion, prioritize traffic based on flow class, and maintain high throughput under heavy load.

The launch of Tomahawk 6 is not simply a product announcement—it represents Ethernet’s transformation from a general-purpose networking solution into the backbone of AI infrastructure. With unmatched bandwidth, extremely low latency, open interoperability, and high scalability, Broadcom positions itself as a key enabler between cloud service providers and AI compute platforms.

Broadcom reports that hyperscale data center operators and top-tier OEM partners have shown strong interest in TH6. It is expected to become a cornerstone product for AI network deployments in the coming years. Broadcom also emphasizes its continued investment in CPO ecosystems, AI-aware routing, and reliability enhancements—aiming to lead the Ethernet evolution for the AI era in close collaboration with industry partners.

For Paid Members ,SemiVision will discuss topics on

  • Tomahawk 6 SerDes Architecture

  • Broadcom Tomahawk Series Overview

  • Broadcom’s CPO Technical Strategy

  • Broadcom Tomahawk 6: Defining the Next Era of AI Networking

  • Key Takeaways from Broadcom’s Tomahawk 6: Redefining AI Network Infrastructure

  • Ethernet as the “Neural Backbone” of AI Infrastructure

  • Scale-Up Ethernet Framework (SUE, Specification RM101)

This post is for paid subscribers

Already a paid subscriber? Sign in
© 2025 TSPA Semiconductor
Privacy ∙ Terms ∙ Collection notice
Start writingGet the app
Substack is the home for great culture

Share