Tech Disruptors Scale Out to Scale Up in Broadcom’s AI Network
23 snips
Jul 15, 2025 Ram Velaga, Senior Vice President at Broadcom, leads discussions on the future of AI networking. He highlights the battle between copper and optical technologies, emphasizing the need for open standards. Velaga explores Broadcom's Tomahawk 6 and Ultra innovations, focusing on their potential to redefine network infrastructure. He also sheds light on the advantages of Ethernet over proprietary systems like NVLink, advocating for vendor neutrality. The conversation reveals how simplicity, bandwidth, and new deployment strategies will shape the next generation of AI networks.
AI Snips
Chapters
Transcript
Episode notes
Ethernet Dominates Scale-Out AI Networks
- AI workloads operate as massive distributed systems requiring high-performance networking.
- Ethernet's openness and performance have made it the dominant choice for scale-out AI networks.
Ethernet Fits Scale-Up Networking
- Scale-up AI networking connects tens to hundreds of XPUs within a rack for memory sharing.
- It's simpler than scale-out, making Ethernet an ideal, open solution for scale-up as well.
Tomahawk 6 Powers High Bandwidth
- Future XPUs will have massive HBM bandwidth, requiring higher network bandwidth for memory sharing.
- Tomahawk 6 switch provides 100 terabits bandwidth, supporting both scale-up and scale-out scenarios effectively.