Marvell Technology has launched a new network switch capable of moving data at 102.4 terabits per second, a chip it says is built specifically for the inside of AI data centers. The company announced the switch on Monday, positioning it as a way to speed up the flow of information between the thousands of processors that train and run large AI models. The move puts Marvell in a direct battle with Broadcom, which has dominated the market for high-speed data center switches in recent years.
What the new switch does
The switch handles 102.4 Tbps of bandwidth, enough to connect hundreds of GPUs or other AI accelerators inside a single cluster. Marvell designed the silicon to handle the unique traffic patterns of AI workloads, where massive bursts of data move between compute nodes during training. The company claims the chip can cut the time it takes to shuffle data around a cluster, which in turn can reduce the overall time needed to train a model. Lower latency and higher throughput mean fewer idle GPUs, a major cost factor for operators of large AI data centers.
Marvell’s announcement comes as demand for AI infrastructure continues to surge. Cloud providers and large tech firms are building out clusters that require faster and more efficient networking to avoid bottlenecks. The switch supports the latest Ethernet standards, making it compatible with existing data center gear.
Competition with Broadcom
The new switch enters a market where Broadcom’s Tomahawk series has been the incumbent choice for many hyperscale data centers. Broadcom’s current top-end switch also offers 102.4 Tbps, so Marvell is matching it on raw bandwidth. But Marvell is betting its architecture will perform better on real-world AI workloads. The company hasn’t disclosed pricing, but analysts expect aggressive pricing to win design wins from cloud customers looking for an alternative to Broadcom.
Broadcom has long held a strong position in merchant silicon for networking, and it has deep relationships with the biggest data center operators. Marvell will need to convince those same companies that its switch offers enough of an advantage to switch suppliers. The stakes are high: data center networking silicon is a multi-billion-dollar market, and AI is driving the next wave of growth.
Why the timing matters
Marvell chose to launch now as AI model sizes continue to balloon. GPT-4 and its successors require clusters of tens of thousands of GPUs, and that number is only expected to grow. Network switches are a critical piece of the puzzle because a slow network can leave expensive compute silicon idle. Marvell’s switch is designed to be a direct replacement for existing gear, so operators can upgrade without redesigning their entire network fabric.
The company also noted that the switch supports the emerging Ultra Ethernet Consortium standard, a specification backed by several tech giants seeking an open alternative to InfiniBand. That could help Marvell win business from customers who want to avoid vendor lock-in.
Broadcom has not yet responded to Marvell’s announcement. The two companies have faced off before in other networking segments, but the AI data center market represents the most lucrative battleground yet. Marvell’s switch is expected to begin sampling to customers later this quarter.
