Fast and Optimal Mapping for Accelerator Modeling and Evaluation (arxiv.org)

A recent announcement in the realm of hardware architecture introduces the Turbo-Charged Mapper (TCM), a groundbreaking tool designed to optimize the mapping process for accelerators used in deep neural network (DNN) computations. The significance of TCM lies in its ability to ensure optimal mappings—crucial for accurately evaluating and designing accelerators—by employing a novel concept called dataplacement. This innovative approach allows for a more efficient analysis and comparison of mappings, leading to a dramatic reduction in search space by up to 32 orders of magnitude. What sets TCM apart from previous methods is its capability to perform full mapspace searches and deliver optimal results in under a minute, making it the first mapper to achieve this level of efficiency in a feasible runtime. In contrast, earlier techniques often relied on heuristic approaches that resulted in energy-delay products up to 21% higher than optimal, even when given substantially more time. This advancement not only enhances the performance evaluation of hardware designs but also significantly impacts the development of more efficient and powerful AI accelerators, potentially transforming the landscape of AI/ML hardware optimization.