Copper's reach is shrinking so Broadcom is strapping optics directly to GPUs

Broadcom is advancing its optical interconnect technology to enhance the performance of AI clusters by integrating optics directly with GPUs. Traditional copper connections are becoming inadequate due to their limited range and bandwidth, especially as data rates exceed 200 Gbit/sec. Broadcom's VP, Manish Mehta, highlighted that copper can only effectively transmit signals for about three to five meters before degradation occurs. To address this, Broadcom is exploring co-packaged optics (CPO), which combines optical engines with GPUs, significantly increasing interconnect bandwidth to approximately 1.6 TB/sec. This innovation could enable large-scale systems with hundreds of GPUs to function as a single unit, contrasting with current scale-out systems that rely on slower Ethernet or InfiniBand networks. Broadcom's experiments have demonstrated promising results, achieving error-free performance with a test chip designed to emulate a GPU. This approach not only enhances bandwidth but also reduces power consumption per port. While other companies like Intel and AMD are also investigating similar technologies, Broadcom's focus on CPO could position it as a leader in the next generation of AI computing infrastructure.

- Broadcom is integrating optical interconnects directly with GPUs to improve AI cluster performance.

- Traditional copper connections are limited in range and bandwidth, prompting the shift to optical solutions.

- Co-packaged optics (CPO) could allow hundreds of GPUs to operate as a single system.

- Broadcom's tests show significant increases in bandwidth and reductions in power consumption.

- Other tech companies are exploring similar optical technologies, indicating a competitive landscape.