Does Broadcom compete directly with POET with their "silicon photonics (SiPh)"?
Exclusive: Broadcom exec weighs in on Nvidia's NVLink strategy and silicon photonics future
Jay Liu, Taipei; Charlene Chen, DIGITIMES Asia
Monday 4 August 2025
The Open Compute Project APAC Summit (OCP APAC) will officially open in Taipei in early August 2025. American IC design giant Broadcom, which holds a critical position in cloud AI application-specific integrated circuits (ASICs) and networking chips, is one of the event's key exhibitors.
Ahead of OCP APAC, DIGITIMES conducted an exclusive interview with Broadcom senior vice president and general manager of the Core Switching Group, Ram Velaga.
Velaga discussed Broadcom's vision for the cloud ecosystem, challenges in switch chip development, and the hot market interest in technologies such as silicon photonics (SiPh) and Ethernet. Crucially, he also clarified Broadcom's current stance toward Nvidia's NVLink camp.
The bandwidth bottleneck challenge
How does Broadcom solve the main challenges currently facing developing cloud AI switch chips?
Velaga stated that the key challenge lies in diminishing returns from process node improvements, which people have been calling "the end of Moore's law."
However, cloud computing demands ever-faster networking chips. As Meta stated a couple of years ago, sometimes their GPUs are sitting idle for close to 60% of the time, waiting for the data to traverse the network.
Every time the switch bandwidth is doubled, 80% to 90% is reduced in power and cost. This indicates that transmission speed is not only the largest bottleneck in cloud computing today, but also a key factor that customers want to optimize to lower deployment costs.
Given the limited benefits from process scaling and tightening physical constraints like power consumption and chip size, achieving bandwidth doubling has been Broadcom's focus in recent years.
Besides cultivating strong technical capabilities, recruiting top talent, and working closely with customers to meet exact specification needs, it is even more important to build "parallel teams" developing various new technologies and products simultaneously.
This approach shortens the product cycle from sampling to mass production to within a year and allows experimentation across different technology upgrade paths tailored to diverse customer requirements.
Silicon photonics adoption timeline
.
What is Broadcom's view on the adoption pace and competitive landscape of SiPh?
Velaga stated that any new technology has a gestation period.
Customers are unfamiliar with potential issues arising during system integration, so helping customers identify and resolve these problems while building trust is the greatest challenge.
SiPh is no exception. Broadcom understands that widespread adoption requires accumulation over "multiple generations," following a long flat growth curve until a breakthrough moment triggers rapid penetration.
Preliminarily, Broadcom estimates it will take at least three to four years before SiPh products reach significant mass production scale and achieve 5% to 10% market penetration.
As a pioneer in SiPh, Broadcom had been shipping SiPh solutions to customers for field trials three to four years before Nvidia announced its offering. Consequently, they feel "very, very good" about their long-term development in this field.
However, Broadcom does not see SiPh as a winner-takes-all market that will generate massive revenue. Instead, it should be regarded as an "enabling technology." The core goal is to empower the entire cloud computing ecosystem to freely develop their own compute chips and systems based on SiPh.
If SiPh and compute chip products can be developed more easily, the AI ecosystem will flourish—that is Broadcom's ultimate objective.
The Ethernet versus NVLink debate
What is the perspective on Nvidia opening up NVLink IP? Will Broadcom cooperate or go its own way?
NVLink's so-called openness is not open, says Velaga.
Nvidia aims to increase NVLink's penetration to pull more cloud computing into its closed ecosystem rather than genuinely creating an open AI ecosystem.
Nvidia GPUs will only recognize their own ConnectX series interface cards or NVLink switches. So regardless of marketing language, any product developed using NVLink IP must remain tied to Nvidia's ecosystem to access full functionality.
The only truly open ecosystem in the world is Ethernet.
What Broadcom builds is less important than the ultimate goal: to create a fully open ecosystem that drives industry-wide technological advancement.
For example, Broadcom helped found the Ultra Ethernet Consortium (UEC) three years ago. Its mission is to prove that large-scale-out networks do not require InfiniBand; Ethernet alone suffices.
The sole technical bottleneck lies in RDMA specifications. Once overcome, whether Broadcom produces the Ethernet chips becomes less relevant. With UEC standards established, anyone can build their own data transmission chips and systems according to the new specifications without needing to purchase licenses or sign private agreements like with NVLink.
All IP and technology are open and freely usable. Broadcom believes nobody in the AI market wants a company to monopolize influence or dictate what others can or cannot do after controlling certain specifications.
Recently, AMD launched its MI series products, whose transport infrastructure is also based on Ethernet. Therefore, Broadcom expects Ethernet to be the mainstream technology for cloud AI data transmission expansion over the next three to four years.
The best aspect of Ethernet chips is that switches and switch chips can come from any supplier without concerns over proprietary technology exclusivity. Broadcom is confident Ethernet technology will have greater room for growth in cloud computing going forward.
Article edited by Jerry Chen
digitimes.com
 |
