Broadcom's Stock Soars as Tomahawk 6 Chip Tackles AI Bottlenecks

Broadcom Inc. is making strategic strides to capture a larger share of the burgeoning AI infrastructure market. Recently, the semiconductor giant unveiled its most advanced data center switch chip to date, the Tomahawk 6, which has commenced mass production.

This chip is pivotal as it addresses performance bottlenecks related to communication speed between GPUs—often the limiting factor when supercomputers with over 100,000 GPUs are in operation. Ram Velaga, Broadcom's Senior Vice President of the Core Switching Group, stated that a single Tomahawk 6 chip can handle the workload of six older-generation chips, enabling faster AI training and more efficient inference calculations while reducing infrastructure redundancy.

The real challenge lies in the current scenario where even NVIDIA's top-tier GPUs spend 60%-70% of their time idle—akin to sports cars stuck in traffic, waiting for network communication to match computing speed. Broadcom's Tomahawk 6 aims to alleviate these "data bottlenecks." Although specific client names are undisclosed, early adopters, including leading cloud service and network equipment companies, have begun deploying these chips to build massive GPU clusters.

Despite the Tomahawk 6’s high price tag, nearly double that of its predecessor, Broadcom is confident its performance enhancements justify the premium. Velaga confirmed that individual chips are priced below $20,000, with discounts available for bulk purchases. The product is set for full release in July, positioning Broadcom as a key player in next-generation AI infrastructure amidst the current drive to optimize power usage, connectivity, and expenditure in AI data centers.

Broadcom has also announced the delivery of the Tomahawk 6 switch series, boasting a single-chip switching capacity of 102.4 Tbps—double the bandwidth of existing Ethernet switches. With its novel scalability, energy efficiency, and AI-optimized functionalities, Tomahawk 6 is tailored for scalable AI networks, offering unparalleled flexibility through support for 100G/200G SerDes and co-packaged optical modules.

This switch chip marks a breakthrough by enabling a single chip to support 1,024 100G SerDes, facilitating the deployment of extended copper-distance AI clusters and optimizing the usage of native 100G interfaces on XPUs and optical devices. Its co-packaged optics option offers the lowest power consumption and latency, enhancing system reliability.

Overall, Broadcom's innovations in the AI semiconductor realm reflect strong market demand, driven by extensive technical expertise and a solid customer base. This strategic focus positions Broadcom optimally for leveraging the ongoing AI technology advancements and maximizing long-term business growth.