The iteration budget is the binding constraint in heterogeneous multi-chiplet design space exploration, not the search algorithm. Each simulator evaluation for a chiplet configuration generates detailed execution timelines, resource utilization breakdowns, memory access patterns, and communication behavior reports. Beacon is a multi-agent framework from USTC that reads those reports directly, identifies which chiplet, which layer, and which parameter is the bottleneck, and makes targeted adjustments. The contrast with data-driven DSE is sharp: data-driven methods train on final PPA metrics and need many evaluations to implicitly reconstruct the parameter-objective relationship. Beacon reads the intermediate reports and gets that relationship for free.
The architecture is three-tier: a model-level agent identifies which DL model operator is saturating resources, a layer-level agent pinpoints which layer within that operator is the constraint, and a solution agent proposes parameter changes grounded in what the reports actually show. RAG memory lets agents recall prior configurations and their outcomes. The framework uses a ReAct loop (reasoning + acting) against a simulator adapter, so the same agent logic works across chiplet evaluation frameworks without retraining.
Chiplet DSE already runs longer than monolithic DSE because heterogeneous configurations multiply the search space. Any method that gets better results in fewer simulator calls reduces the actual calendar time from architecture decision to verified configuration. Vendors selling homogeneous chiplet toolkits with fixed interconnect assumptions lose ground to frameworks like this that treat the entire heterogeneous parameter space as searchable. Expect this pattern to appear in commercial DSE tools within 12-18 months.