“Broadcom’s Subtle Expansion into AI Infrastructure” • The Register

GPUs play a crucial role in AI infrastructure, but interconnect fabrics are equally vital for training large models. Broadcom has been increasingly influential in developing these interconnects, ranging from scale-out Ethernet fabrics to chip-level communications. Unlike Nvidia, Broadcom offers merchant silicon, enabling diverse companies, like Google and Apple, to utilize its technology without widespread acknowledgment. As GPU deployments grow, the demand for efficient switches has surged; Broadcom’s high radix Tomahawk 6 switches can connect large clusters of GPUs with fewer units. The company is also pushing for Ethernet as a backbone for rack-scale architectures, further supporting connectivity with its co-packaged optics (CPO) technology. By moving optics onto the switch package, they aim to enhance efficiency. Additionally, Broadcom is advancing chiplet technology for multi-die architectures, focusing on higher interconnect speeds. Its innovative approaches and designs, expected to enter production by 2026, highlight its role in the evolving AI landscape.

Source link