Abhinav Kothiala

Summary

Daniel is joined by Abhinav Kothiala, a principal product manager for the Synopsys Ethernet IP portfolio. He has over 12 years of experience across engineering and product management, spanning SoC design, functional verification, and building wireless connectivity platforms and IoT products. He also holds two patents in… Read More

Key takeaways

• AI training workloads require lossless, deterministic networking because tail latency from a single delayed packet can stall an entire accelerator cluster, making conventional best-effort Ethernet insufficient for scale-up fabrics.
• UALink and ESUM represent complementary approaches to open multi-vendor scale-up networking—UALink enables memory-semantic, tightly coupled accelerator communication, while ESUM extends Ethernet for lossless scale-up with smoother integration into existing infrastructure.
• Future scale-up fabrics must evolve across four dimensions: higher per-lane speeds (toward 400Gbps), wider domains with higher-radix switches, longer reach via optical interconnects, and smarter in-network compute (INC) to reduce redundant data movement.

Why this matters

As AI clusters scale to thousands of accelerators, the network fabric has become a first-class architectural decision—not background infrastructure—directly determining training efficiency, hardware utilization, and total cost of ownership in next-generation data centers.

Entities

Show notes

Daniel is joined by Abhinav Kothiala, a principal product manager for the Synopsys Ethernet IP portfolio. He has over 12 years of experience across engineering and product management, spanning SoC design, functional verification, and building wireless connectivity platforms and IoT products. He also holds two patents in… Read More

Themes

• ai