Weimin Fu
"This dataset records every trial from a study comparing large language models with classical and learned optimizers on microarchitecture design space exploration, for two RISC-V cores. CVA6 is a six-stage in-order core with 13 configuration knobs and 46,656 valid configurations; BOOM is an out-of-order core explored in both a 12-knob and a 25-knob space. Each trial pairs a configuration with measurements from the flow design teams actually use: cycles per instruction from cycle-accurate RTL simulation, and total cell area, maximum frequency, dynamic power, and leakage power from logic synthesis against a 32-nm standard cell library. The CVA6 database holds 42,565 trials from 11 language models and eight baselines across six workloads and five seeds, under prompt conditions that vary whether knob names are real or obfuscated and whether functional descriptions are supplied. Of the 36,085 model-driven trials, 32,284 carry the natural-language rationale the model gave for its proposal, averaging 612 characters, which makes the dataset usable for work on reasoning quality and not only on search efficiency."