What does an AI Workload actually need from silicon? imec's cross stack view

As AI workloads increasingly dictate hardware requirements, the semiconductor industry needs a common, vendor-neutral language to connect process and packaging decisions to real-world AI performance. This talk introduces imec's AIStack intelligence: a benchmarking platform and layered framework that maps AI workloads — from inference to training — down through system architecture, device integration, and ultimately to process and materials choices such as lithography parameters and node scaling. Drawing on early results, including a GPU total-cost-of-ownership analysis, we will show how fab-level decisions ripple through to workload economics, and outline imec's ambition to build an open ecosystem where hardware and AI software communities can jointly optimize the stack.