- Built an energy-efficient multicore system to tackle the dark silicon problem using asymetric & reconfigurable CPU cores that operate by turning on/off lanes in sections of the superscalar out-of-order pipeline.
- Formulated it as an optimization problem that involves mapping applications onto different core types & reconfiguringthe cores, to maximize the system throughput while operating under a power budget.
- Developed techniques to tackle the challenging combined problem of thread-to-core mapping & core reconfiguration in just a few milliseconds using machine learning and global optimization algorithms.
- Our approach outperforms prior work, namely core-level gating and Flicker (technique aimed at homogeneous reconfigurable multicores) by up to 30% and 15% respectively.