Americas

  • United States

Broadcom’s new switching chip links GPUs, aims to boost AI networks

News Analysis
Apr 25, 20232 mins
Data CenterNetwork Switches

The new Jericho chip can connect up to 32,000 GPUs concurrently and promises shorter job completion times for AI workloads.

asia night connect network shutterstock 1111879247
Credit: Shutterstock

Broadcom’s new networking chip, called the Jericho3-AI, is designed to connect supercomputers and features a high-performance fabric for artificial intelligence (AI) environments.

Broadcom has three switch families: the high-bandwidth Tomahawk switch platform, which is used primarily within data centers; the lower bandwidth Trident platform, which offers greater programmability and deeper buffers, making it more suited for the edge; and the Jericho line, which sits somewhere between the other two and is best suited for low latency interconnects.

Jericho3-AI is targeted at AI and machine-learning backend networks where the switch fabric handles spraying of traffic on all network links and reordering of that traffic before delivering to the endpoints. It also has built-in congestion management capabilities for load balancing and minimizing network congestion. 

The Jericho3-AI has a top throughput of 28.8Tb/s. It has 144x SerDes lanes operating at 106Gbps PAM4 and supports up to 18 800GbE/36 400GbE/72 200GbE network-facing ports.

Beyond speeds and feeds, Jericho3-AI features improved load balancing over the prior generation to ensure maximum network utilization and congestion-free operation. No packet jitter and a high radix (meaning support for a large number of ports in the switch) allow Jericho3-AI to connect to 32,000 GPUs collectively.

The Jericho3-AI fabric is designed to lower the time spent moving data around during AI training and inference jobs. AI training can take weeks if not months, and it requires an awful lot of data to be moved around the network.

Broadcom claims that because of its performance improvements, the Jericho3-AI chip reduces the cost of running AI workloads, thus making it so valuable that it effectively pays for itself.

“The benchmark for AI networking is reducing the time and effort it takes to complete the training and inference of large-scale AI models. Jericho3-AI delivers significant reduction in job completion time compared to any other alternative in the market,” Ram Velaga, senior vice president and general manager of the core switching group at Broadcom, said in a statement.

Jericho3-AI is now available to qualified customers.