Broadcom Retargets Jericho at AI Clusters

Broadcom is challenging the conventional topology linking AI engines in training supercomputers/clusters, offering a version of its Jericho Ethernet switch chip set for AI.

Joseph Byrne

Breaking the convention that hot technologies belong to a single era, Broadcom is repositioning its Jericho network processor, which was conceived during the turn-of-the-century telecom boom, for AI systems. The Jericho3-AI (BCM88890) upgrades the Jericho2c+ chip Broadcom sampled in 2020, doubling the serdes speed to 106 Gbps (raw) to support 800GbE front-panel ports. Despite increasing bandwidth, it reduces features not critical for building large-scale AI-training systems.

Its companion cell-based fabric chip Ramon3 quintuples throughput from its predecessor’s 9.6 Tbps to 51.2 Tbps. Harnessing multiple Jericho3-AI and Ramon3 ICs, designers can create systems with 32,768 ports running at 800 Gbps. Sampling Jericho3-AI and Ramon3 now, Broadcom expects the 5 nm chips to qualify for production in a year. Together these chips can implement an interconnect for AI-training supercomputers and clusters. Adapting the end-to-end, credit-based scheduling for moving packets across a router backplane to a distributed leaf-spine switch, the chips can balance networking loads, handle traffic congestion, and mitigate the effects of link failures.

High-radix switching, massive bandwidth, and end-to-end scheduling intersect AI in the massive systems cloud-service providers build to train large models. Google’s TPUv4-based systems, first deployed in 2020, employ up to 4,096 AI accelerators. Newer systems will employ more nodes to accommodate bigger models such as GPT-4. Whereas Google, like high-performance computing companies, employs a torus topology and proprietary interconnect, Broadcom anticipates some companies’ AI-training systems will connect AI-accelerator nodes with switches that allow nonadjacent nodes to exchange data without passing through other nodes.