Software Challenges in Heterogeneous Computing: A Multiple Case Study in Industry

One way to improve the performance of embedded systems is through heterogeneous platforms, i.e., using hardware containing more than one type of processor, like CPU + GPU or CPU + FPGA. This approach has shown improved performance, particularly in the domain of artificial intelligence, in which computationally demanding models must be trained and executed. However, these computational environments pose various challenges to software engineering, since applications must be designed and developed differently while accounting for target hardware architectures that are inherently different. Companies interested in migrating to heterogeneous platforms must be aware of the changes to the software development processes that are required to accommodate such solution. In this paper, we conducted a multiple case study that aims to discover the challenges and common practices when developing software for heterogeneous platforms in industrial contexts. First, we organized semi-structured interviews with companies on the automotive, automation, and telecommunication domains. Then, we analyzed and structured the data in order to create a classification describing the software engineering challenges faced by companies using heterogeneous platforms. The companies involved in this study are in different stages of maturity concerning the use of heterogeneous platforms. Thus, the classification takes into consideration these levels to describe the challenges accordingly.