AI’s High Speed Chase in Networking

Chaos Lever Podcast

Ned and Chris discuss the evolution of data center networking technologies, particularly focusing on the rivalry between Ethernet and InfiniBand in the context of AI and high-performance computing.

Ethernet vs. InfiniBand:

This episode takes listeners into the world of data center networking, as Ned and Chris dissect the critical role of Ethernet and InfiniBand technologies in AI's rapid evolution. They explore how these networking giants influence the performance and efficiency of AI workloads and high-performance computing, highlighting their implications for future tech innovations.

Links:

• Intel’s Emerald Rapids XCC model will have 61 Billion transistors across 64 cores
• Seagate’s Exos Mosaic hard drive will have 30TB of capacity
• Nimbus Data has a 100TB SSD
• Broadcom’s Tomahawk 5 chip does 51.2 Tbps at line rate
• https://www.simplilearn.com/tutorials/machine-learning-tutorial/what-is-epoch-in-machine-learning
• https://community.fs.com/article/need-for-speed-%E2%80%93-infiniband-network-bandwidth-evolution.html
• https://www.arubanetworks.com/faq/what-is-spine-leaf-architecture/
• https://avinetworks.com/glossary/equal-cost-multi-path-routing-ecmp/
• https://en.wikipedia.org/wiki/Go-Back-N_ARQ
• There are five big improvements the UEC hopes to make with their v1.0 of the standard
• https://engineering.fb.com/2024/03/12/data-center-engineering/building-metas-genai-infrastructure/