InfiniBand is the dominant method of connecting AI processing systems. Is there some kind of significant performance benefit? Or is it more of a legacy choice? There are other options on the horizon but you need to understand the differences and how they need to be built to take advantage of needed performance.
Ethernet is the future of AI networking but it’s going to take cooperation between the vendors to help it overcome the performance of InfiniBand. Is it really Ethernet? Or something different running over Cat 6 cables? How will the Ultra Ethernet Consortium make peace and pave the way for Ethernet to assume the leadership role in AI networking?
In this episode I discuss the dominance of InfiniBand as well as the way Ethernet is changing to support large latency-sensitive flows in the modern AI networking world. We look at the companies leading the way in the Ethernet world and how the formation of the UEC is the first step to unseating InfiniBand.
Links
- Wikipedia InfiniBand Page
- Infiniband vs. Ethernet
- Another InfiniBand Comparison to Ethernet
- Next Platform’s Analysis of the InfiniBand Ethernet discussion
- Broadcom’s Take on Ethernet for AI
- Broadcom on Ethernet Fabrics for AI
- Next Platform Looks at Broadcom Jericho3 AI
- DriveNets on the Need for Ethernet for AI Networking
- Next Platform Looks at Cisco G200
