It's been some time since I've written about Marvell. The semiconductor company continues to innovate, and Marvell executives recently briefed my firm's chief analyst Patrick Moorhead and me regarding its recent announcement. I want to share my takeaways from that discussion.
AI is driving the need for data center headroom
Marvell focuses on silicon for connectivity infrastructure across the data center, carriers, automotive applications and enterprise networking. Our discussion in early March centered squarely on the data center and three key market trends: artificial intelligence (AI) applications, network bottlenecks and bandwidth growth. For the first trend, Marvell rightly focuses on the larger workloads driven by AI. It’s easy to see the importance of this, given the world’s recent fascination with ChatGPT, Bard and Bing AI, drawing millions of users to use AI applications for the first time.
Interestingly, AI is also triggering the second trend toward network bottlenecks, given the need for lower latency in AI applications. Vast amounts of data are thrown into AI models, both for training the algorithms and for computational processing. In that context, ultra-low latency has become a requirement for processing things quickly and getting the most out of AI. Finally, the general bandwidth growth driven by cloudification is a third evident trend creating a need for additional headroom for future applications and workloads.
Marvell targets speed and savings with Nova and Teralynx 10
To address these market trends and challenges, Marvell has announced two new pieces of silicon: Nova and Teralynx 10. Marvell is positioning Nova as the industry's first 1.6T PAM4 DSP, and the underlying architecture is jaw-dropping. It includes eight 200G lanes at half the typical module count. This results in a 30% savings in cost and the same in power reduction per bit. The company says that Nova also improves reliability by a factor of two and, most importantly, supports a multi-vendor ecosystem. Moorhead and I jokingly asked executives for the punchline: with all this good news, where’s the catch? They assured us that there is no catch—just smart engineering. Still, all joking aside, Nova is well positioned to provide data center operators with what they need, given the apparently unstoppable upward trends in data proliferation and consumption.
Teralynx 10 is an ultra-low-latency 51.2T switch that aims to support advanced telemetry at an astounding 80% cost savings. The Teralynx architecture enjoys a significant install base, given that Marvell has deployed it at the huge scale of five million 400G ports shipped to date. The Teralynx 10 optimizes AI, machine learning (ML) and data center network fabrics. Marvell also touts the benefits of using Nova and Teralynx 10 in tandem; it says that in deploying both, data center infrastructure providers can leverage common firmware, software and link training schemes to speed deployment and achieve a faster time to market. On the surface, it’s hard to argue with its assertions.
I would also be remiss if I didn’t mention another Marvell article penned by Moorhead that speaks to the company’s recent announcement with Amazon Web Services (AWS). The company supports the hyperscaler with electro-optics, networking, security and storage improvements. If interested, you can find it here.
I continue to be impressed with Marvell's achievements in supporting connectivity infrastructure. As a former marketer, I always gauged success for such engineering-intensive products with the mantra of "better, faster, cheaper." Marvell is hitting on all three vectors with its recent announcement and, in the process, enabling a broad ecosystem of high-performant networking infrastructure. I’ll be eager to see what Marvell comes up with next.