Modern FPGA placement techniques with hardware acceleration

Date

2019-09-13

Authors

Dhar, Shounak

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

In deep sub-micron technology nodes, Application-Specific Integrated Circuits (ASICs) are becoming expensive to design and manufacture. For this reason, Field Programmable Gate Arrays (FPGAs), which are general purpose and flexible programmable hardware, are gaining more design wins in low volume and fast evolving applications. Modern FPGAs are becoming popular in high performance data analytics, search engines, autonomous cars, communication and networking applications. FPGAs are also accompanied with a complete Computer-Aided Design (CAD) toolchain, that is used to optimally map and fit the design applications or workloads onto the underlying target FPGA device. These design applications mapped onto the FPGA demand high maximum achievable clock frequency (Fmax) and low power consumption while maintaining a low compilation time, which is a major hindrance in widespread adoption of FPGAs. The focus of this Ph.D. dissertation is the placement problem for FPGAs, which takes a major portion of the FPGA CAD tool runtime. A new algorithm for spreading cells during FPGA global placement is proposed, which achieves better wirelength and routing congestion and takes less runtime than the algorithm used in the state-of-the-art academic FPGA placer. We also propose FPGA acceleration of various subsystems of an analytic global placement algorithm, including wirelength gradient computation and spreading, which achieves significant speedup over the multi-threaded CPU version. A new detailed placement algorithm is proposed, which offers better tradeoff between quality and runtime compared to existing methods. This algorithm is also accelerated on a GPU and an FPGA, achieving significant speedup over multi-threaded CPU implementation. Another detailed placement algorithm is also proposed which physically re-aligns timing critical paths and improves Fmax with minimal runtime overhead. Both of these algorithms for detailed placement have shown good results on industrial benchmarks and have been integrated into an industrial FPGA CAD tool flow

Description

LCSH Subject Headings

Citation