Broadcom (AVGO, Financial) is making strategic moves to capture a larger share of the booming AI infrastructure market. The company has commenced mass production of its most powerful data center switch chip to date, the Tomahawk 6. This chip addresses a critical bottleneck in AI supercomputers equipped with over 100,000 GPUs, which lies not in the computational chips but in the communication speed between GPUs.
According to Ram Velaga, Broadcom's Senior Vice President of Core Switching, a single Tomahawk 6 chip can handle the workload of six previous-generation chips. This advancement allows for faster AI training and more efficient inference operations without increasing the number of GPUs, thereby reducing infrastructure redundancy.
Currently, even top-tier GPUs from companies like NVIDIA experience 60%-70% idle time due to communication delays, akin to sports cars stuck in traffic. Broadcom's new chip aims to alleviate these "data bottlenecks." Early adopters, including leading cloud service providers and network equipment companies, are deploying this chip to build the world's largest GPU clusters.
Despite its high price—nearly double that of its predecessor—Broadcom believes the enhanced performance justifies the premium. The Tomahawk 6, priced under $20,000 per chip, is set for full market release in July and is expected to be a vital component in next-generation AI infrastructure.
