- Cisco unveiled its answer to the scale across problem data centers are facing
- Its new P200 chip and 8223 router support 51.2 Tbps
- The system relies on deep buffering, setting up the market for technology bifurcation
Nvidia may have popularized the term ‘scale across,’ but the tech giant clearly isn’t the only one who’s been thinking about the new kind of networking muscle data centers will need in the AI era. Cisco has come out swinging with a new AI networking system it thinks will rise to the challenge as data centers look to build ever-larger GPU clusters.
The system includes Cisco’s Silicon One P200 chip as well as its 8223 router and supports a whopping 51.2 terabits per second for intense AI workloads. Cisco provided all sorts of fun stats on power efficiency, the system’s compact form factor and security, which you can check out here. But the true attention grabber is its secret sauce: deep buffering.
If you’re confused because you remember buffering being a bad thing back in the ancient days of music and video streaming, let us explain.
The problem that Cisco (and Nvidia, for that matter) is trying to solve is how to send massive amounts of data (that is, AI traffic) over long distances between different data centers. The trick is they need to do it without dropping packets, since packet loss can cause AI workloads (especially training) to fail and require starting over. But avoiding dropped packets is easier said than done.
Guru Shenoy, SVP for Cisco’s Provider Connectivity Group, told Fierce that when data is sent over long lengths, the data flow can be quite bursty. If the chip on the end of the line isn’t able to hold the data contained in that burst when it comes through, that results in packet loss. Buffers, he said, are like buckets that can hold large amounts of data and help smooth the traffic flow.
Thus, deep buffering means deep buckets and less packet loss.
But not everyone agrees that this is the right approach. Nvidia is notably not using deep buffering with its Spectrum-XGS Ethernet platform due to concerns around added latency. Instead, Dell’Oro Group VP Sameh Boujelbene noted, it’s going with something called Auto-Adjusted Distance Congestion Control and using end-to-end telemetry to solve the same problem Cisco is tackling.
Rakesh Chopra, Cisco SVP and Fellow for Silicon One, acknowledged there’s an impression in the market that deep buffering hurts AI performance, though he argued this isn’t true.
“We’re here to say it’s actually an ‘and.’ What you really need is the intelligent congestion control that some people are talking about, but that is not enough to move this data between multiple sites,” Chopra said. “The only way to not drop those packets on the floor during failure conditions is to have these deep buffers.”
So, what does this mean for data centers?
Boujelbene told Fierce the scale across market (which she described as data center interconnect for AI data centers), is at a fork in the road.
“I expect to see bifurcation in the market, just like what we’ve seen inside the AI data centers,” she concluded.
Don’t wait, enter Fierce Network's Innovation Awards before it's too late! The final deadline for entries is on October 17. Use discount code DIANA50 to save $50 on your entry. Click here for more info and enter today.