Broadcom Releases Chips to Interconnect AI Supercomputers – Technology

Reuters – Broadcom (AVGO.O) on Tuesday unveiled a new chip for interconnecting supercomputers for artificial intelligence (AI) work using networking technology already widely used.

Broadcom is a major supplier of chips for Ethernet switches, the primary way computers are interconnected in traditional data centers.

However, the rise of AI applications such as OpenAI’s ChatGPT and Alphabet Inc’s (GOOGL.O)’s Bard pose new challenges for networks in data centers. To answer questions with human-like answers, such systems must be trained using vast amounts of data.

That job is too big for a single computer chip to handle. Instead, you need to split your job across thousands of chips called graphics processing units (GPUs). These chips act like one giant computer and must do work for weeks or months at a time. So the speed at which individual chips can communicate is important.

Broadcom announced on Tuesday a new chip, the Jericho3-AI, capable of connecting up to 32,000 GPU chips. The Jericho3-AI chip competes with another supercomputer networking technology called InfiniBand.

Currently, the largest manufacturer of InfiniBand equipment is Nvidia Corp (NVDA.O), which acquired InfiniBand leader Mellanox for $6.9 billion in 2019.

Nvidia is also the market leader in GPUs. The Nvidia-Mellanox system is one of the fastest supercomputers in the world, but many companies have decided to give up on the Ethernet marketed by various companies and buy both their GPUs and network equipment from the same supplier. Reluctant, says Ram Velaga. Senior Vice President and General Manager of Broadcom’s Core Switching Group.

“Ethernet is available from multiple vendors. There is a lot of competition,” Velaga said. “Even if we don’t make the best Ethernet switches, someone else will. InfiniBand is a unique, single source, vertically integrated solution.”