Abstract:
Approximate hardware designs are largely acquired via automated frameworks, most of which rely on iterative search to find feasible approximate instances in the design sp...Show MoreMetadata
Abstract:
Approximate hardware designs are largely acquired via automated frameworks, most of which rely on iterative search to find feasible approximate instances in the design space. However, the combinatorial explosion restricts the exploration to only a subset of design space and requires heuristic pruning. As a result, many feasible designs are discarded. Although search algorithms like Monte Carlo Tree Search (MCTS) provide a reasonable balance between exploring new designs and exploiting good designs while always looking at the global tree representation, they tend to consume large runtimes. Recently, machine learning models have been utilized to predict the output error of approximate designs during the search in an effort to reduce the runtime. While suitable for rather small sized circuits, accuracy degradation affects the quality of the design space exploration for large practical designs from complex domains e.g., machine learning. In this paper, we introduce static and dynamic checkpoints, which are carefully marked points within the search space, and propose DeepApprox, a framework that combines the checkpoints with light-weight deep learning quality estimators to compensate the error induced by predictors. As a result, the search algorithm can timely adjust paths taken during the search. The results demonstrate an improved yet scalable design space exploration with hardware area reduction of up to 52.6 % on real-work benchmarks in comparison with the most competitive state-of-the-art frameworks.
Date of Conference: 01-03 July 2024
Date Added to IEEE Xplore: 25 September 2024
ISBN Information: