Broadcom has recently unveiled its latest generation chip, Sian2, specifically designed to support high-performance optical networks behind artificial intelligence clusters. The Sian2 module reportedly doubles the bandwidth of its predecessor while integrating multiple reliability features to ensure data accuracy during high-speed transmission.

In environments running large language models (LLMs), multiple servers typically collaborate, with each server handling a portion of the model. These distributed model segments need to regularly exchange data to coordinate tasks, necessitating shared network connections between servers. Data center operators often employ fiber optic technology to build such networks, as fiber optics offer significantly faster transmission speeds than traditional copper wires, making them ideal for the rapid exchange of vast amounts of data in AI clusters.

Within AI servers, graphics processing units represent data as electrical signals, but to send this data to another server in the cluster via a fiber optic network, it must first be converted into optical signals. This conversion is handled by specialized network devices called transceivers, and Broadcom's newly launched Sian2 chip powers these data center transceivers.

Sian2 is a digital signal processor whose integrated circuits are optimized to efficiently convert electrical signal data into optical signals and perform the reverse operation—converting received optical signals back into electrical signals for server processing. In optical networks, data is transmitted through different "channels," each handling an independent data stream. The Sian2 chip increases the data transmission rate per channel to 200 gigabits per second, doubling the rate of its predecessor, Sian.

The increased bandwidth means that the number of transceivers required to build a fiber optic network is halved, reducing hardware procurement costs and decreasing power consumption, thereby further lowering operational expenses. Additionally, Sian2 utilizes a 5-nanometer manufacturing process and integrates laser driver components, which are critical for generating the optical signals needed for fiber optic transmission.

When optical signals are transmitted through a fiber optic network, scattering can cause errors that accumulate and affect the processing of AI models. To address this, Sian2 supports Forward Error Correction (FEC) technology, which improves network reliability by transmitting multiple copies of data. Receiving servers can compare multiple data copies, selecting consistent and accurate data for processing or using intelligent algorithms to estimate correct data in the event of errors.

Vijay Janapaty, Vice President and General Manager of Broadcom's Physical Layer Products division, stated, "A digital signal processor with 200 gigabits per channel is essential for building high-speed optical links in the next-generation AI infrastructure."

Currently, Broadcom has provided samples of the Sian2 chip to early customers.