Skip to Main content Skip to Navigation
Theses

High-level area-performance-energy estimation framework for FPGA-based accelerators

Abstract : In recent years, the complexity of system-on-chip (SoC) designs has been dramatically increased. As a result, the increased demands for high performance and minimal power/area costs for embedded streaming applications need to find new emerged architectures. The trend towards FPGA-based accelerators is giving a great potential of computational power and performance required for diverse applications. The advantages of such architectures result from many sources. The most important advantage stems from more efficient adaptation to the various application needs. In fact, many compute-intensive applications demand different levels of processing capabilities and energy consumption trade-offs which may be satisfied by using FPGA-based accelerators. Current researches in performance, area and power analysis rely on register-transfer level (RTL) based synthesis flows to produce accurate estimates. However, complex hardware programming model (Verilog or VHDL) makes FPGA development a time-consuming process even as the time-to-market constraints continue to tighten. Such techniques not only require advanced hardware expertise and time but are also difficult to use, making large design space exploration and time-to-market costly. High-Level Synthesis (HLS) technology has been emerged in the last few years as a solution to address these problems and managing design complexity at a more abstract level. This technique aims to bridge the gap between the traditional RTL design process and the ever-increasing complexity of applications. The important advantage of HLS tools is the ability to automatically generate RTL implementations from high-level specifications (e.g., C/C++/SystemC). The HLS tools provide various optimization pragmas such as loop unrolling, loop pipelining, dataflow, array partitioning, etc. Unfortunately, the large design space resulting from the various combinations of pragmas makes exhaustive design space exploration prohibitively time-consuming with HLS tools. In addition, to thoroughly evaluate such architectures, designers must perform large design space exploration to understand the tradeoffs across the entire system, which is currently infeasible due to the lack of a fast simulation infrastructure for FPGA-based accelerators. Hence, there is a clear need for a pre-RTL and high-level framework to enable rapid design space exploration for FPGA-based accelerators.
Document type :
Theses
Complete list of metadata

https://hal-uphf.archives-ouvertes.fr/tel-03366300
Contributor : Marie Zoia Connect in order to contact the contributor
Submitted on : Tuesday, October 5, 2021 - 3:58:01 PM
Last modification on : Wednesday, October 6, 2021 - 3:36:15 AM

Identifiers

  • HAL Id : tel-03366300, version 1

Collections

Citation

Mariem Makni. High-level area-performance-energy estimation framework for FPGA-based accelerators. Computer Science [cs]. Université de Valenciennes et du Hainaut-Cambrésis; Ecole nationale des ingénieurs de Sfax (ENIS), 2018. English. ⟨NNT : 2018VALE0042⟩. ⟨tel-03366300⟩

Share

Metrics

Record views

10