Ciso: New router for distributed AI workloads
With its new 8223 router and the Silicon One P200 network processor, Cisco wants to make it easier to build scalable AI networks across multiple data centres.
(Image: Cisco)
- Arne Bauer
Network equipment provider Cisco has introduced the 8223 model, a router specially designed for distributed AI workloads. The router delivers a total throughput of 51.2 Tbit/s and is designed to eliminate bottlenecks in so-called scale-across architectures, such as those that occur when several AI data centers are linked.
Videos by heise
According to the manufacturer, the 8223 offers switch-like energy efficiency in a compact design. The router offers 64 ports of 800G each on three height units and processes up to 20 billion packets per second. It also allows 800G Coherent Optics for data center interconnects (DCI) and metro connections (metro DCI) up to 1000 kilometres.
New Silicon One P200 as the centrepiece
At the heart of the platform is the specially developed Silicon One P200 network processor, which, according to Cisco, achieves its maximum aggregation bandwidth of over 3 exabits per second in a 3-layer topology. The chip's deep buffering function is designed to cushion strong traffic fluctuations, for example during AI training, and thus prevent network outages. The P200 is programmable to support future protocols without a hardware upgrade.
At launch, the router will run Cisco's open-source network operating system SONiC, with support for IOS XR planned. In the future, the chip will also be used in modular platforms, disaggregated chassis, and Nexus systems. The security features include line-rate encryption with post-quantum-resistant algorithms and continuous monitoring. The first systems have already been delivered to hyperscalers.
(mki)
