Marvell Technology said it has demonstrated the industry’s first 3nm data infrastructure silicon to boost data transfer speeds up to 45% faster than existing approaches with multichip products.
Its announcement on Wednesday comes one day after Broadcom announced delivery to some customers of its Jericho3-AI for Ethernet in a 32,000 GPU cluster.
While the two products are in different categories, they show a lightning-fast move by silicon providers to make products for cloud companies and large data centers to handle acceleration, often across artificial intelligence networks.
“Networking large systems of chips together is one of the biggest bottlenecks for increasingly larger generative AI models, “ Dylan Patel, an analyst at SemiAnalysis, told Fierce Electronics. IDC has predicted global spending on AI will grow 27% by 2026, reaching $300 billion or more.
The high-speed interconnects from Marvell use the SerDes moniker and are produced on TSMC’s 3nm process, the company announced Wednesday. Marvel had earlier produced the first data infrastructure SerDes silicon based on TSMC’s 5nm process.
The new technology will be combined with 2.5 D and 3D packaging for complex semiconductors and promises to reduce the needed number of pins, traces and circuit board space, thus reducing cost. A hyperscaler data center could have tens of thousands of SerDes links where a cost reduction even in pennies would be of value.
SerDes is using SerDes and interconnect technologies in Octeon processors and Brightlane automotive Ethernet chipsets along with other silicon including custom ASICs. Moving to 3nm allows engineers to lower costs and power consumption as well, while maintaining signal integrity and performance, the company said.
“Marvell’s successful production of 3nm SerDes and interconnets marks the latest step in helping cloud service providers to stay ahead of the ever-escalating demand for higher speeds and more traffic,” said Alan Weckel, co-founder of research firm 650 Group in a statement.
“Marvell has consistently stayed ahead on SerDes technologies and that looks to continue with their 3nm offerings that will have an advantage in time to market, power and performance,” Patel added. “With that said, competition is heating up from Broadcom, AlphaWave and Synopsys trying to catch up or even surpass Marvell, but Marvell likely keeps its lead to some extent with the next generation technologies.”
Broadcom said Jericho3-AI will scale connectivity to 32,000 GPUs, each with 800 Gbps in a single cluster. It relies on 5nm technology.
“In a world where companies such as Nvidia are increasingly turning to proprietary fabrics for HPC and AI, Broadcom is looking to push Ethernet with higher speeds and more sophisticated packet processing to keep the open standards competitive with closed standards,” Patel said.
RELATED: Generative AI is driving a gold rush for performance tools