Design space exploration is a critical development stage for hardware design. With the growing complexity of applications and the corresponding hardware we design to accelerate them, the task of exploring the wide variety of potential hardware architectures is challenging and time-consuming. In this domain, hardware designers must use heuristics to intelligently explore the design space. Even in small designs there may be hundreds or thousands of possible design points to evaluate, and it is rarely feasible to evaluate all possible design points. In this paper, the authors propose an extension to design space exploration that not only explores the design space, but also automatically generates the design models used to perform the design space exploration.

In automating the step of model creation through the use of an L₁-regularized least squares regression technique, designers now have the ability to add design parameters and automatically infer which parameters influence the quality of output designs. In this paper, there is both a description of the model generation technique as well as a case study using two computer vision algorithms. Although the computer vision algorithms are simple, well-studied algorithms, and the parameters explored are also comparatively simple, the authors present a compelling vision of a system where both the generation of design quality models and design space exploration using those models are automated.

The studied design parameters are intuitive parameters that have a clearly expected influence on one or more measurement metrics. In this sense, the results of automated model generation are unsurprising, as the parameters are expected to have these effects. However, the authors demonstrate the excellent performance of selected designs, as well as a technique that opens the door to automatic model generation and design space exploration of much larger design spaces.

This initial paper has significant promise, as this could fill a critical need in model generation and support efficient design space exploration. Although further study is needed to ensure that this technique remains feasible and scalable as we scale the size of designs, the number of explored design parameters, and the number of interacting design blocks under simultaneous exploration, this paper presents a vision of automated model generation and exploration that takes an important step toward producing high-quality hardware implementations without significant pre-characterization or designer guidance during the exploration process.