Training modern Large Language Models (LLMs) requires tens of thousands of accelerators acting as a single "AI Supercomputer." To build this infrastructure, we must first address the CPU bottlenecks of traditional general-purpose networking. This talk analyzes why standard TCP/IP processing limits training performance and introduces the role of "Kernel Bypass" and SmartNICs in offloading network processing from the host CPU. We will explore the transition toward hardware offloads like RDMA to achieve the throughput and nanosecond-latency required for accelerator-to-accelerator communication. The session concludes with the architectural challenges of running lossless transport over lossy Ethernet and the design patterns required to ensure reliable delivery in massive hyperscale environments.
Sujithra Periasamy is a Senior Software Engineer at Google with extensive experience in software design and development, specializing in networking, distributed systems, and telecommunications. She is recognized for her strong problem-solving abilities and capacity to quickly understand complex systems and adopt new technologies. Currently based in the San Francisco Bay Area, she has been with Google since January 2017.
