This article is more than 1 year old

Cisco's latest switch packs 32 800G ports into a pizza box

If you need even more speed, Nvidia and Broadcom have silicon twice as fast

Cisco this week unveiled one of its highest performance and lowest form factor switches yet – a diminutive 1RU pizza box packed with 32 800Gbit/sec ports – at the Open Compute Platform Summit in San Jose.

Do you need 800Gbit/sec networking? Probably not – unless you're at webscale, hyperscale, or your enterprise is really gung-ho on the metaverse or AI/ML. These are exactly the kinds of customers Cisco is targeting with its Nexus 9232E and Cisco 8111 switches.

According to Guru Shenoy, VP of product management for Cisco's Mass-Scale Infrastructure Routing Group, the switches are actually the same in terms of hardware. The Nexus-branded appliance runs Cisco's in-house network operating system, while the other is designed for disaggregated cloud environments using something like SONiC.

The appliances are powered by Cisco's Silicon One G100 processor. The 7nm switch ASIC is capable of 25.6Tbit/sec of bandwidth across 256 100Gbit/sec serializer/deserializers. That bandwidth feeds 32 800Gbit/sec QSFP-DD-800 ports, which in turns can be broken out into 64 ports at 400Gbit/sec or – you guessed it – 256 ports at 100Gbit/sec.

Alongside the new switches, Cisco also announced a pair of 800Gbit/sec optical pluggables. The first employs eight 100Gbit/sec links, while the second condenses that down to a pair of 400Gbit/sec links. However, we're told Cisco plans to offer a single stream 800Gbit/sec pluggable when the technology has been standardized.

As you might expect, Cisco is positioning the appliances as aggregation or spine switches, used to network multiple racks of systems in 400Gbit/sec or 800Gbit/sec links. While this might sound like a lot of bandwidth to cram into a single system, Cisco argues the approach allows it to achieve far better efficiencies.

Workloads like 5G, IoT, edge compute, AI/ML, and 8K video streaming not only require greater network densities, but the ability to serve these applications in a power efficient manner, Shenoy said. "Sustainability is top-of-mind for everybody, and that's what these platforms enable us to do."

According to Cisco, the appliances consume 77 percent less power for the same bandwidth compared to using 12.8Tbit/sec switches. Cisco is able to make this claim because you would need roughly six 12.8Tbit/sec switches operating in a mesh to achieve the same bandwidth. By Cisco's estimate that works out to be about 3000W.

Of course, by this logic a 51.2Tbit/sec switch would be even better. As it happens, Broadcom – which has had 25.6Tbit/sec switch silicon, albeit using 50Gbit/sec SerDes, on the market for some time now – made the same 6x efficiency claims for its Tomahawk 5 ASICs when they began sampling to customers in August.

And Broadcom isn't alone. Nvidia also has a 51.2Tbit/sec switch in the works, which the chipmaker claims will begin sampling to customers later this year.

Pressed on this, Shenoy argued that Cisco's rivals may have faster silicon out to ODM customers, but they're still a ways off putting switches with 100Gbit/sec SerDes in customers' hands. "I do believe the silicon is available, but I don't believe there is any device that has them that is shipping," he said of Broadcom and Nvidia.

"We believe that the large majority of the market will not be ready right now for 51.2Tbit/sec," he added.

How many customers there are, desperate to add a bunch of 800Gbit/sec ports to their networks, remains to be seen. But a Dell'Oro report from earlier this year suggests it's probably not many. It put market penetration at roughly ten percent of shipments — indicating that even 400Gbit/sec networking is still in its infancy. ®

More about


Send us news

Other stories you might like