Recently, Cisco held a Tech Talk focused on Cisco’s Silicon One and how the company believes you can converge a network without compromise. This means simplifying operations, maintaining security, and optimizing performance. Typically, companies have to choose between one of the three and compromise between the other two. The discussion also shed light on how Cisco is dealing with the rise of AI.
Pierre Ferragu, who heads up the global technology infrastructure research team at New Street Research, hosted the session. He was joined by Eyal Dagan, EVP of the company’s Common Hardware Group, and Rakesh Chopra, Cisco Fellow in the Common Hardware Group.
The need to do something different
Rakesh started the discussion by looking back to the start of Cisco Silicon One in late 2019. “We realized that we at Cisco—and everybody else in the industry—had been making the same mistakes repeatedly,” he said. You will get the same outcome if you approach a problem with the same organizational structure and technology.”
The company shook things up. “We first created a new organization at Cisco that Eyal runs,” he said. “This new organization is focused on building one architecture in silicon that you can use across your network and also across different business models.”
Covering the full network and solution space
Rakesh added that Cisco focused broadly on the entire network and solution space. The company invested many years and dollars to enable the convergence of routing and switching and plowed over a billion dollars into Silicon One. Cisco sees this as a fundamental industry shift.
“We are the industry’s first truly scalable networking silicon architecture,” he said. “And that becomes very important when you start thinking about the role of silicon in AI networking.”
Some industry watchers have called out Cisco for being proprietary because of Silicon One, but there is a precedent behind what Cisco is doing. General purpose silicon is just that – general purpose. This makes it OK at many things but not great at anything. The best example of this is the GPU. A CPU does a lot of things well, but not accelerated computing, which is why there is so much investment today in GPUs for AI. Fortinet also makes its silicon to handle security processing. Cisco requires customer silicon for networking and chooses to build it themselves to deliver the performance and features it knows its customers need today.
Helping customers build AI networks on Ethernet
Rakesh likes to think about AI in two buckets. He sees the first bucket—using artificial intelligence to improve Cisco products and services—as a large part of Cisco’s revenue. However, as important as that is, the company's main focus is selling Cisco products to enable its customers to build AI networks.
Regarding web scalers, Eyal noted that two kinds of networks in data centers are critical to running AI apps. In addition to the front-end network we’re all familiar with, there’s a back-end network, typically InfiniBand, that has historically been used to connect storage clusters and the like. Cisco also sees Ethernet as a solution, especially in the world of web-scale.
“We have customers who are deployed at scale with Ethernet-based AI networks,” he said. “And all of the others are actively trialing Ethernet AI.” Cisco also says it can provide significant efficiency boosts for power-hungry, saving a megawatt of power for a single AI/ML cluster.
Choosing the right silicon model
Eyal came on to discuss costs and operating in the web-scale world. Three silicon models exist: merchant, ASIC, and fabless COT (or customer-owned tooling). “Cisco used to be an ASIC house, and we still use, in some cases, merchant silicon,” he said. “But we moved in a big way in the last five or six years to a fabless COT.”
At the same time, there has been a move from branded boxes to white boxes in the back end. This is for a simple reason—the bill of materials. White boxes, in the right environment, can be much more affordable.
Final thoughts
Since launching Silicon One, Cisco has been working hard to develop it into a viable alternative to the typical ASIC approach. With the rise of AI and its massive impact on networking, the company has pivoted swiftly to enable customers to utilize its converged networks to run taxing AI apps.
The presentation by Eyal and Rakesh provided a helpful glimpse into their strategic thinking and was a welcome respite from the AI vaporware we’re exposed to daily.
Zeus Kerravala is the founder and principal analyst with ZK Research.
Read his other Network Computing articles here.
Related articles: