Broadcom is at the forefront of efforts to address the bandwidth limitations for artificial intelligence (AI) by directly integrating optical connectivity into GPUs. According to TechSpot, Broadcom’s innovative silicon photonics technology for GPUs is set to revolutionize data connection capabilities in the context of increasing bandwidth demands for AI tasks. The company’s “optical engine” can provide up to 1.6 TB/s of connection bandwidth, offering significantly higher speeds than traditional copper connections, while also reducing power consumption.
At the recent Hot Chips conference, Broadcom revealed its work on integrating optical interconnect components into GPUs using synchronous optical packaging (CPO) technology. This technology, made possible through TSMC’s CoWoS packaging technology, allows optical transceivers to be placed directly on the GPU. Manish Mehta, Vice President of Broadcom’s optical systems division, highlighted that CPO technology can connect up to 512 GPUs across 8 racks, promising a highly capable unified computing system for AI tasks.
While this cutting-edge technology has not yet been integrated into commercial products, Broadcom’s advancements open up a promising future for data connectivity in the realm of AI.