Data-Efficient Cross-Generation Adaptation for Microcontroller Performance Screening

Microcontroller (MCU) performance screening requires accurate estimation of the maximum operating frequency (Fmax) under worst-case voltage and temperature conditions, but early production stages provide only a limited number of labeled characterization samples. Historical measurements from previous MCU generations are often available, yet their reuse is hindered by cross-generation shifts in Speed Monitor (SMON) features and target-frequency distributions. This paper presents an empirical study of dataset-level adaptation protocols for low-data Fmax prediction across related MCU generations. Two incremental training schemes are compared: mixed incremental training, where legacy and target samples are interleaved during model updates, and baseline-first adaptation, where legacy datasets initialize the predictor and target samples are progressively introduced. Per-product normalization and target-prioritized sample weighting are used to improve cross-dataset compatibility. Experiments on a target MCU product with three legacy industrial datasets show that naive mixed training is unstable in the low-label regime, whereas baseline-first adaptation yields smoother convergence and improved target-domain sample efficiency. Results further indicate that, in this setting, training protocol and source-target data handling have a larger impact than increased model complexity.