Beyond the Network: Building a Tiered Fabric for Agentic AI Inference – Marvell Technologies Triple-T
Agentic AI transforms inference into a distributed, memory-intensive systems challenge — where KV cache state must be placed at the right performance tier with predictable latency and bandwidth. No single technology solves this alone. This session presents a systems architecture spanning on-package HBM, host-attached memory, a shared CXL-based KV-cache tier, and durable storage — unified by a Teralynx Ethernet fabric delivering the high-throughput east-west connectivity required for agent coordination, cache movement, and multi-rack inference. Attendees leave with a practical framework for designing AI infrastructure as an integrated compute, network, and memory system.
Ashwin Gopalan is Director of Solutions Architecture for AI Fabrics at Marvell Technology, where he leads the Solutions Architecture for Marvell’s AI datacenter portfolio — spanning scale-out switching (Teralynx T100/T200), scale-up interconnects, memory fabric, and custom silicon solutions. He works at the intersection of silicon innovation and real-world AI infrastructure deployment, translating high-fidelity simulation and lab validation into compelling customer engagements at conferences including OFC, Computex, and Hot Chips.
Prior to Marvell, Ashwin spent nearly eleven years at Apple as a Senior Network Engineer within Global Networking Services, where he designed and deployed some of Apple’s most critical network infrastructure
He is passionate about bridging the gap between pre-silicon simulation and production AI infrastructure — and helping customers navigate the rapidly evolving landscape of AI networking, memory fabrics, and custom accelerators.
