Computer architecture design and research is often inefficient and ad hoc due to the significant costs of hardware simulators. We must urgently address these costs as technology scaling presents greater challenges in design complexity, energy efficiency, and system integration. I present the case for statistical inference in architectural design, enabling holistic solutions that (1) control complexity using inference, (2) extract efficiency using hardware specialization, and (3) analyze interaction within integrated systems using modular models. Throughout, inferential models act as surrogates for simulators and capture the complexity of simulated architectures with the speed of analytical equations. This speed transforms the way architects reason about design priorities: energy efficiency, microarchitectural adaptivity, chip multiprocessors, and multiprocessor heterogeneity.
Sunday, March 29, 2009
Free and open to the public