Broadcom is out with another high-speed networking ASIC. This time, it is a high-speed router chip targeting the AI build-out with the Broadcom Jericho4 generation. Built on a 3nm process and with 200G PAM4 signaling, this is a new generation of parts. Notably, the introduction of 3.2Tbps HyperPorts enables high-speed transmission between locations.
Broadcom Jericho4 51.2Tbps AI Router Chip Now Shipping
The Jericho4 line is not going to be found in most enterprise networks. Instead, this ASIC is designed for massive AI clusters, the ones targeting 1 million AI accelerators and featuring things like line rate MACsec encryption.
Perhaps the most notable feature is the new 3.2Tbps HyperPorts for those that want massive bandwidth tied to a port when 4x 800GbE is not enough.
Broadcom is also talking about their capabilities with a 100Km+ RoCE deployment. This is important for site-to-site communication, especially as large AI clusters colocate next to power sources.
We actually started showing some ways these long-distance links for AI need to be bridged in the future in our Going 800Gbps at up to 1000km with the Marvell COLORZ 800 piece earlier this year.
Jericho4 is a 51.2T switch chip. Broadcom is not calling this a UEC switch, but said it is “fully compliant with specifications developed by the Ultra Ethernet Consortium (UEC)” (Source: Broadcom.)
The new switch chips are available now.
Final Words
It feels like Broadcom has a massive amount of switch/ router silicon lines now. Between Broadcom Tomahawk 6 for high-throughput, Broadcom Tomahawk Ultra for HPC and AI, and Jericho4 for AI, and Broadcom Trident 5, just to name a few lines. Broadcom has this diagram to describe where Tomahawk Ultra, Tomahawk 6 and Jericho4 fit.
It sees Jericho4 as running in the higher-levels of the scale-out networking tiers, especially those between data centers and across large campus deployments.
Overall, another neat Broadcom switch ASIC announcement.
