Moritz Joseph and Ananda Samajdar presented their paper "AIRCHITECT: Learning Custom Architecture Design and Mapping Space" at the Design, Automation and Test in Europe (DATE) conference in Antwerpen.
The paper proposes a novel approach to custom architecture design and mapping space exploration, which typically involves costly iterations of simulations and heuristic tools. Instead, we investigate the possibility of using machine learning to learn the optimization task and predict optimal design and mapping parameters for custom architectures, bypassing the need for exploration.
Through three case studies, our paper demonstrates that it is possible to capture the design space and train a model to generalize predictions for optimal configurations when queried with workload and design constraints. They formulate the architecture design and mapping as a machine learning problem and design and train a custom network architecture called AIRCHITECT. It is capable of learning the architecture design space with as high as 94.3% test accuracy and predicting optimal configurations that achieve on average 99.9% of the best possible performance on a test dataset with GEMM workloads.
The proposed approach has significant implications for custom architecture design and mapping space exploration, potentially leading to faster and more efficient design and deployment of custom architectures with maximum performance and energy efficiency.