In edge AI inference, FPGA Overlay Accelerators offer superior performance-area balance but often lack accurate maximum frequency prediction and efficient Design Space Exploration (DSE) due to complex parameterization. To address these limitations, we propose a Bayesian optimization-based DSE framework using the Tree-Structured Parzen Estimator. We introduce dependency-aware design space modeling (DAM) to prune the parameter space and a custom regression predictor (CREP) for frequency. To accommodate varying computational constraints, we present two implementations: a lightweight framework, DEFA, and an advanced framework, DEFA-ML. DEFA saves computational cost of CREP and exploration runtime for rapid optimization. Conversely, DEFA-ML leverages a complex prediction model and a Large Language Model assistant for superior optimization quality. We validate the framework on the Altera FPGA AI Suite. For frequency prediction, DEFA provides fast and conservative predictions, while DEFA-ML achieves high linearity with a mean absolute percentage error of 4.21%. In throughput optimization, compared with the FPGA AI Suite baseline optimizer, DEFA improves throughput by 30.16% and DEFA-ML improves by 35.12%. In area-throughput tradeoff optimization, compared with the baseline, DEFA improves by 39.05% and DEFA-ML improves by 43.11% for a single target model. For multiple target models, DEFA improves by 40.76% and DEFA-ML improves by 47.66%.
Transistor-level design is pivotal for the accurate evaluation of FPGA architectures. Since traditional linear models are increasingly inadequate in advanced technology nodes, simulation-driven approaches have become the standard for FPGA architecture exploration. However, due to the non-analytical nature of delay measurements from simulations, the transistor sizing process becomes a black-box optimization problem. COFFE2 [1], [2], the state-of-the-art academic sizing tool employs a division-based brute-force approach, which is relatively time-consuming and may lose optimal solutions. In this paper, we propose a two-phase transistor sizing methodology and enhance the COFFE2 framework with Bayesian Optimization, which is well-suited for black-box optimization problems. Our proposed approach not only achieves a 11.7% improvement in the quality of results but also reduces runtime by ~40%-70%. Through extensive benchmarking on a complete design flow, from transistor-level sizing to routing with VTR benchmarks, we demonstrate that FPGA architectures optimized by our approach offer an 11.0% reduction in the area-delay product, proving the efficacy of our method.