Broadcom Stretches Switching Across Datacenters With Jericho 4

Speaking metaphorically, the blast area of a datacenter should only be so large, for mechanical as well as technical reasons, in order to mitigate the risk of a failure taking out too many machines. This is one reason why companies build datacenter regions, with multiple datacenters, rather than one giant datacenter.

This has been true of cloud datacenters running CPU systems for more than a decade now, and it is now becoming true for AI datacenters as companies try to build supercomputers that span not tens of thousands of compute engines, but rather hundreds of thousands to millions of GPUs and XPUs that occasionally need to be used all at once to train a large model.

For a long time, Cisco Systems, Juniper Networks (now part of Hewlett Packard Enterprise), and Huawei Technologies sold datacenter operators big wonking routers with custom ASICs to interconnect their datacenters, but the hyperscalers and cloud builders balked at the cost such behemoths and convinced merchant switch chip makers like Dune Networks (acquired by Broadcom in November 2009 for $178 million) and Broadcom to start adding routing functions to their switch chips. Because these hyperscalers and cloud builders had created their own Linux-based network operating systems for both switching and routing, they could immediately make use of these routing functions and start weening themselves off Cisco and Juniper kit.

The fun bit is when these switch/routing hybrid chips, which have deep buffers on their packet processors so they can deal with congestion better and not drop packets to deal with that congestion, were chosen for selected parts of the interconnects used by hyperscalers and cloud builders in other parts of their networks where high latency and dropped packets were a no-no.

This is where the “Jericho” StrataDNX switch/router chips and their related “Ramon” fabric elements come in, which are an important part of the Broadcom datacenter networking portfolio. While we were away on vacation at a lovely lake in northern Michigan last week, Broadcom rolled out the Jericho 4 switch/router chip and the related Ramon 4 fabric element, boosting the capacity of the datacenter interconnects and deep packet switching jobs that it can tackle – and importantly, positioning Jericho 4 as the means to interlink datacenter halls crammed with GPUs and XPUs into clusters with 1 million compute engines sharing work.

The Jericho 4 chip completes the set of networking ASICs that Broadcom is revealing this year and that are now sampling to equipment makers for what is shaping up to be a revolutionary year in AI compute and networking in 2026.

The Tomahawk Ultra chip announced in July is at one end of the spectrum, and is designed to be used for scale up networking to link the memories of GPUs and XPUs so they can share work more easily and quickly as well as a replacement for low latency, high bandwidth InfiniBand switch fabrics commonly used in traditional HPC simulation and modeling clusters as well as the back end networks for lashing together GPU nodes together. The Tomahawk 6 chips announced in June (there are two of them) can be used for scale up interconnects for GPUs and XPUs (and offer more aggregate bandwidth and therefore higher radix to do so), but are really designed for the scale out networks on the back end of AI clusters (and certain HPC clusters) as well as the front end networks that are used for all manner of distributed systems and across systems.

As we pointed out above, the Jericho 4 chip is aimed at scale out networks that have big data flows and are sensitive to congestion as well as for interconnecting datacenters and scaling AI applications across datacenters. These jobs require bigtime bandwidth, and Jericho 4 has some tricks to bring a lot more bandwidth to bear than has been available to date as well as integrated security for those who are sensitive about data moving across datacenter boundaries.

The big innovation with Jericho 4 is something Broadcom calls a HyperPort, which is a kind of aggregation of four 800 Gb/sec ports that provides 3.2 Tb/sec of bandwidth and that actually looks like a single port to load balancing software running on the Jericho 4 switch/routers. The idea is that you can reduce the number of ports to deliver any given amount of bandwidth across the datacenters by a factor of four.

“We greatly reduce the number of ports, so the chance that you can load hashing, collisions, or polarization will be greatly reduced,” Hasan Siraj, head of software products and ecosystem at Broadcom, tells The Next Platform, adding that this HyperPort approach is better than traditional ECMP link aggregation because load balancing and congestion control software sees it as one port. As a result, a HyperPort can drive its component links at 70 percent higher utilization rates in the datacenter interconnect use case than prior approaches.

Like the Tomahawk 6 and the Tomahawk Ultra chips, the Jericho 4 is etched using 3 nanometer processes from Taiwan Semiconductor Manufacturing Co and also employs a chiplet design as you can see from the feature image of the Jericho 4 chip shown at the top of this story. If you look at the feature image and the other images for Jericho 4 in this story, you will see that there are two versions of Jericho 4. One variant has all Ethernet interface for fixed form factor switches, the other is around half Ethernet interface and around half fabric interface (the precise breakdown is not clear) for chassis and distributed scheduled fabric applications. Both Jericho 4 versions support HBM and use the same Condor SerDes chiplets. To our eye, it looks like there is a 25.6 Tb/sec all Ethernet version with 128 lanes running at 200 Gb/sec and a 51.2 Tb/sec hybrid version that has 128 lanes running at 200 Gb/sec for linking to Ramon 4 fabric elements and 128 lanes running at 200 Gb/sec for Ethernet.

Jericho 4 represents a big jump in capacity for the Jericho ASIC line. The Jericho 2 chip that was announced way back in March 2018 started production in February 2019, and it had 9.6 Tb/sec of aggregate bandwidth to drive 24 ports at 400 Gb/sec. The Jericho 2 also had 8 GB of HBM memory for its packet buffer, which allowed it to have more buffer capacity than could have been added to the ASIC itself and the very fastest speeds available.

With the Jericho 3-AI launched in April 2023, Broadcom took the same “Peregrine” SerDes used in the Tomahawk 5 switch ASICs that came out in August 2022 and created a device with 51.2 Tb/sec of aggregate bandwidth that could scale out to 32,000 GPUs in a single network using the Jericho 3-AIs in the top of racks and a leaf/spine network of Ramon 3 fabric elements to link them all together. (We do not know how much HBM memory is on this Jericho 3-AI device, which was etched using 5 nanometer processes from TSMC.)

This was pretty good scalability, but the pipes were not all that fat. With Jericho 4, the pipes are getting fatter and slightly more of them are being ganged up to drive a larger number of devices within a datacenter or links across datacenters. Like this:

We do not know how much HBM is on the Jericho 4, but we do know that the bandwidth is 2X that of the buffer memory on the Jericho 3-AI.

What is interesting is that a network of 4,500 Jericho 4 switches can provide 36,000 HyperPorts running at 3.2 Tb/sec. First, this would make a killer back end network for an AI cluster, although we suspect it would be a whole lot more expensive than a network based on Tomahawk 6 switches. And second, by using the 51.2 Tb/sec Jericho 4 chip and its HyperPorts, you can get 4.5X the bandwidth across a datacenter interconnect. It will take 4.5X the number of Jericho 4 chips to do this – 4,500 versus 1,000 – but at least it is possible to gang up the ports and really boost the bandwidth across the interconnect.

There are a lot of ways to deploy the Jericho 4 as well as the Tomahawk 6 and Tomahawk Ultra chips in AI networks. Here are two scenarios across two interlinked datacenters:

Broadcom suggests that at the very least, for companies that want to deploy 1 million GPUs or XPUs across datacenters, they should use Tomahawk 6 in the leaf/spine network spanning racks and use Jericho 4 to interlink the datacenters, which have 250,000 GPUs or XPUs each.

The Jericho 4 can drive datacenter interconnects that are 100 kilometers or more in length. The Jericho 4 chips are using the same “Condor” Serdes used in the Tomahawk 6 switch chips, by the way. The Jericho 4 chip also sports integrated MACsec encryption (first made popular on campus networks) that can secure data at line rate as it is zipping into and out of the Jericho 4 device.

As with Tomahawk 6, it looks like there are two variants of the Jericho 4: One that uses 100 Gb/sec SerDes and one that uses 200 Gb/sec SerDes. These two SKUs allow customers to dial up the bandwidth and radix they need for their networks.

The Jericho 4 chips are sampling now and are expected to be in switches and routers in production in the first quarter of 2026, says Siraj.