Skip to content

WaytoICT

通往ICT之路 · 光通信与AI网络技术解读

中文English
登录注册
Menu
Technology Optical comms · silicon photonics · CPO → Market Insights Financials · trends · industry → In-Depth Analysis Optical chips, devices & modules → Video Understand a topic in minutes → Documents Latest docs · summit materials →
Menu

A 51.2Tb/s 5nm Monolithic Switch Chip for AI/ML Networks!

Posted on 2025-03-31 by

The Tomahawk 5 (BCM78900 series, abbreviated TH5) chip launched by Broadcom is a breakthrough product designed for next-generation data centers and AI/ML networks. This monolithic switch chip built on a 5nm process sets a new industry benchmark with a 51.2Tb/s switching capacity, doubling the bandwidth over the previous generation, while balancing high performance and low power consumption through multiple technological innovations.

A 51.2Tb/s 5nm monolithic switch chip for AI/ML networks

The core advantage of TH5 stems from its highly integrated architecture. The chip carries 512 106.25Gb/s PAM4 SerDes lanes, supporting flexible port configurations from 100GbE to 800GbE, and dynamically allocates resources through a shared-buffer architecture, significantly improving burst traffic handling capability and quality of service.

A 51.2Tb/s 5nm monolithic switch chip for AI/ML networks

Broadcom’s self-developed Peregrine SerDes technology not only supports multiple physical connection schemes such as direct attach copper (DAC), linear pluggable optics (LPO), and co-packaged optics (CPO), but also achieves signal insertion loss performance of more than 45dB, providing data center operators with cost-effective networking options.

Power control is one of the core challenges in TH5’s design. Through multi-level adaptive voltage scaling (AVS) technology, the chip dynamically optimizes voltage distribution in the range of 0.7V to 0.7875V with 12.5mV precision. Combined with architectural innovations, the typical operating power consumption is ultimately reduced to 450W, further reducing the energy per bit from 12.0pJ of the previous generation to 8.8pJ.

A 51.2Tb/s 5nm monolithic switch chip for AI/ML networks

This achievement benefits not only from the energy efficiency improvement of the 5-nanometer process, but also from the synergistic improvement of microarchitecture optimization and SerDes technology. For example, TH5 adopts a symmetric layout design to reduce the number of physical modules, improves memory utilization through semi-customized packet buffering, and uses feed-through wiring to lower interconnect cost.

In addition, the chip’s power delivery network (PDN) design needs to cope with the extreme operating condition of a 430A current transient within 550 nanoseconds. To this end, the team used a combination of multilayer copper PCB and decoupling capacitors to strictly control voltage ripple within 3%, ensuring system stability.

A 51.2Tb/s 5nm monolithic switch chip for AI/ML networks
)

In terms of packaging technology, TH5 adopts an innovative compressed hexagonal BGA layout (ball pitch 0.9mm) (Fig. 16.1.4), reducing package size while optimizing signal integrity. Through a pre-configured customized test platform, Broadcom completed comprehensive verification of chip-package interaction effects before tape-out, ultimately passing the JEDEC component-level reliability test in one attempt and meeting stringent standards such as high-temperature warpage. In actual deployment, the TH5 validation kit (SVK) achieves efficient heat dissipation with a lidless package combined with an air cooling system (bottom of the figure above), supporting 64 800G ports interconnected through stacked OSPP connectors, showcasing a minimalist system design philosophy.

For optical module integration, TH5 supports co-packaged optics (CPO) solutions, achieving energy efficiency of 4.8pJ/bit (Fig. 16.1.5), saving more than 60% energy compared with traditional fully-modulated optical modules.

A 51.2Tb/s 5nm monolithic switch chip for AI/ML networks

The link performance of the optical modules (as shown below) further verifies their reliability in complex scenarios. The upper-left part of the figure shows the pre-FEC bit error rate (Pre FEC BER) using the PRBS31 test pattern; the upper-right part shows the pre-FEC bit error rate (Pre FEC BER) under real traffic; and the bottom is its module test platform (MTP), used for LRO/LPO optical module validation.

A 51.2Tb/s 5nm monolithic switch chip for AI/ML networks

The left part of the figure below shows that Tomahawk 5 co-packages 8 optical engines, each with a capacity of 6.4T; the right part shows that the Tomahawk 5 co-packaged optical link performance complies with the IEEE 400GBase-FR4 standard.

A 51.2Tb/s 5nm monolithic switch chip for AI/ML networks

From the perspective of technological evolution, the Tomahawk series started at 3.2Tb/s on the 28-nanometer process and, through five generations of iteration, has continuously broken through in process nodes, bandwidth density, and energy efficiency ratio.

Reference: IEEE – Tomahawk 5: 51.2Tb/s 5nm Monolithic Switch Chip for AI/ML Networking.

Together with several bloggers, we have built an ima Q&A tool in the ICT field, currently free to join – welcome to experience it:

A 51.2Tb/s 5nm monolithic switch chip for AI/ML networks

Related Reading

  • Ethernet and Optical Transport – FlexE (2)
  • Introduction to 400G Ethernet Transceivers
  • Flex Ethernet (FlexE) Link Interface Test Method (YD/T 3992-2021)

Enjoy this article?

Subscribe to get new optical communication insights delivered to your inbox. Free, unsubscribe anytime.

Category: Technology

Leave a Reply

You must be logged in to post a comment.

订阅更新

订阅后,新文章发布时第一时间收到邮件通知

李东霏
光互连技术研究者

香农信息技术研究院,专注光模块、CPO、硅光与 AI 光互连的技术解读。

行业报告

光芯片·光器件·光模块 2026 报告汇编(含 OFC 2026)

光芯片 · 光器件 · 光模块
2026 报告汇编(含 OFC 2026)

¥1,900

上/下两册,收录 OFC 2026 一线报告
添加微信 WaytoICT 订购

广告

📢 广告位招商:面向全球光通信客户 侧边栏广告位 Email:Dongfei.li@shannoninstitute.cn

最新文档

  • 互连拐点:11 页解析机架内到多机架 Scale-Up 的 XPO/CPX/OCI 蓝图(OCP 2026 全球峰会)
  • AI 时代的相干光技术:17 页解析 Scale-Across、液冷 XPO 与全谱转发器(OCP 2026 全球峰会)
  • XPO 高密度光互连的系统级评估:20 页解析 212G SerDes 平台与 FRO/LRO/LPO 实测(OCP 2026 全球峰会)
  • CPO 与 OCI 标准及未来技术:17 页 Broadcom 精简版演讲提纲(OCP 2026 全球峰会)
  • VCSEL 高密度互连:7 页解析 50G NRZ「慢而宽」阵列与光 PCIe 演示(OCP 2026 全球峰会)

最新评论

暂无评论,欢迎登录后发表第一条评论。

最近访客

👥 今日访客 -
加载中…
© 2026 WaytoICT | 蜀ICP备2020035321号-2 | 隐私政策 | 免责声明
Wechat Pay
Alipay
请使用 支付宝 或 微信 扫码支付