Transfer Learning in MCU Performance Screening

In safety-critical applications, microcontrollers must meet performance standards, including the maximum operating frequency (Fmax). ML models can estimate Fmax using data from on-chip ring oscillators (ROs). However, when new products are introduced, existing ML models may no longer be suitable and may require updating. Training a new model is challenging due to limited data availability and time needed to acquire Fmax. But data from legacy products, along with pre-trained models, may still be available. We propose using deep-learning models trained on a specific MCU product as feature extractors for new devices, to address the scarcity of labeled data, in a Transfer Learning fashion. Experimental results show that these models can extract useful general features for performance prediction even from new products. As a result, they achieve better performance with significantly less labeled data compared to traditional shallow learning approaches