Highlights:

  • The Sian2 is a digital signal processor, an integrated circuit designed to convert data stored as electrical signals into light.
  • Broadcom’s previous transceiver chip, the Sian, transmits up to 100 gigabits per second per lane, while the Sian2 handles 200 gigabits per second.

Recently, Broadcom Inc. has launched the Sian2, a new chip designed to drive high-speed optical networks that support artificial intelligence clusters.

The company claims that the module offers double the bandwidth of its predecessor. It also incorporates reliability features that prevent errors from infiltrating data as it travels across the network.

A typical large language model operates across multiple servers, with each server hosting a small portion of the LLM. These model fragments must frequently exchange data to synchronize their tasks, which requires the servers to be connected through a shared network.

Data center operators frequently use fiber optic technology to connect their artificial intelligence servers. Since light moves faster through glass than electricity through metal, fiber optic cables offer faster data transmission speeds compared to traditional copper wiring. This makes fiber optics ideal for AI clusters that require high bandwidth, where large amounts of data are constantly exchanged between servers.

The graphics card within an AI server represents data as electrical signals. Before this data can be transmitted over a fiber optic network to another server in the cluster, it must be converted into light. This conversion is handled by a specialized networking device called a transceiver. Broadcom’s newly launched Sian2 chip is specifically designed to power transceivers used in data centers.

The Sian2 is a digital signal processor, an integrated circuit specifically designed to convert data stored as electrical signals into light. It can also perform the opposite function—when a server on an optical network receives data in the form of light pulses, the Sian2 processor transforms the light back into electrical signals that the server can interpret.

Optical networks are structured into lanes, each handling a separate data stream. Broadcom’s previous-generation transceiver chip, the Sian, could transfer up to 100 gigabits of data per second per lane. In contrast, the new Sian2 can reach speeds of 200 gigabits per second per lane.

Doubling the bandwidth per lane reduces the number of transceivers needed to construct a fiber optic network by half. This decrease in hardware reduces procurement costs. Additionally, with fewer chips in the network, power consumption is lowered, leading to further savings.

The Sian2 is produced using a five-nanometer fabrication process. In addition to circuits that convert electrical signals to light and back, it features a component called a laser driver, which plays a crucial role in producing the light used by optical networks for data transmission.

As light beams move through the fiber-optic network of an AI cluster, they scatter within the cables, resulting in data errors. If left uncorrected, these errors can accumulate and disrupt the AI models’ processing. To solve this issue, the Sian2 chip incorporates a widely used error correction technology known as FEC.

FEC enhances network reliability by sending each piece of data multiple times instead of just once. As a result, the server on the receiving end obtains several copies of the data. If all the copies match, the server can determine that no errors occurred during transmission and can continue with processing.

The technology is also beneficial in cases where errors do occur. Typically, errors manifest as discrepancies among the various copies of a data point received by an AI server via a fiber-optic connection. With FEC, the server can make an informed decision about which copies are likely correct and utilize them in its calculations.

Vice President and General Manager of Broadcom’s physical layer products division, Vijay Janapaty, said, “200G/lane DSP is foundational to high-speed optical links for next generation scale-up and scale-out networks in the AI infrastructure.”