High-Level Synthesis for FPGAs-A Hardware Engineer's Perspective

The recent decades have witnessed unprecedented advances in the complexity of digital hardware systems, yet their design methods are still mostly based on manual register-transfer level (mRTL) languages such as VHDL and Verilog, introduced in the 1980s. While allowing exact system description, these languages have low productivity and require special expertise. High-level synthesis (HLS) promises to increase the productivity of hardware design by allowing system description from abstract, timeless source code, which is synthesized into optimized RTL code by an HLS tool according to technological constraints. However, HLS is still seen as somewhat immature technology with a non-consolidated offering of tools with varying features. Furthermore, the quality of results (QoR) of HLS is seen to be worse than with mRTL methods. This study sheds light on the status of HLS today. The emphasis is on field-programmable gate arrays (FPGAs) that allow fast development cycles. The study briefly covers the history of HLS, describes the HLS design flow, and lists the benefits and remaining challenges. The offering of current commercial and academic HLS tools is surveyed along with their features. A literature survey covering academic articles published between 2017 and 2024 on the QoR and productivity of HLS is presented. The results show that a gap of some margin still exists between the QoR of the HLS and mRTL methods. However, in productivity, HLS clearly outcompetes mRTL. Based on the study, several recommendations are made for HLS tool developers to close the QoR gap and accelerate the adoption of the method.