AI/ML Networking Driving Need for Silicon Diversity

Cloud Operators Continue to Push for Silicon Diversity as We Enter the 51.2 Tbps Switch Era

As we engaged with Cloud Operators over the past two quarters, the shift away from COVID-19 supply constraints and toward next-generation build-outs became apparent in a changing of priorities. In our 4Q reports, one could see this with one of the cloud operators’ 800G ramp. What is very clear with 800G, 1.6T, and Ethernet in AI/ML planning is that cloud operators want more vendor diversity at the system and silicon levels. While it may be an evident best practice to be multi-vendor, the COVID-19 supply shock reinforced and made multi-vendor a more critical priority.

If we look at the 3.2 Tbps (100G) generation, Broadcom was the majority of each Cloud operator environment. In our market shares, Broadcom recorded several records in this generation, and almost all cloud traffic ran on Broadcom’s ASICs. With 12.8 Tbps (400G), those same operators began to diversify, with one of the cloud operators moving significant volume over to Innovium’s Teralynx 7. We note the cloud operator also shifted white box suppliers to further its supply chain diversity. 650 Group market share data showed a significant difference in market share between these two generations of ASICs.

At the start of the 51.2Tbps era, we expect cloud operators to increase diversity even further. We see this in our interviews and supply chain studies for both traditional compute and the AI/ML Ethernet networks for back-end and front-end. The four main options that they will be using are Broadcom (Tomahawk 5), Cisco (Silicon One G200), Marvell (Teralynx 10), and Nvidia (Spectrum 4). While additional families and flavors will exist, such as deep buffer ASICs in some AI/ML workloads and DCI, over 90% of the 51.2T volume will be these four silicon offerings over the next 2-3 years.

Silicon and system diversity is why executives at the Broadcom alternative companies enthusiastically talk about proofs-of-concept (POC) and RFP visibility. Cloud operators are giving them more opportunities in the 51.2 Tbps generation to compete for business. A mix of custom OS solutions, traditional vendor NOS, and the SONiC open source OS will run on all these ASICs. With 51.2 Tbps multiple Cloud Operators plan to second source switch silicon.

Given where network designs will go for the rest of the decade, we view vendor diversity and a robust silicon ecosystem as critical to building out AI/ML, pushing the envelope on DCI, and providing an incentive to keep innovating. After all, it won’t be long before 102.4 Tbps announcements begin to hit the market.