The primary constraint in heterogeneous chiplet design is now the computational cost of multi-physics co-simulation, not toolchain capability, particularly for 2.5D and 3D stacks. This simulation demand, critical for AI infrastructure, requires distributing workloads across hundreds of machines or using GPUs, with AI-powered models offering further speed improvements. EDA environments accessible via scripts and APIs are essential for generating the vast data needed to train these AI models, fundamentally changing simulation economics.
The computational burden of chiplet simulation is a growing bottleneck, driving demand for GPU compute and AI-powered EDA tools to accelerate design cycles.