Optimal design space exploration for FPGA-based accelerators: a case study on 1-D FDTD
Date
2015
Authors
Puranik, Mugdha, author
Rajopadhye, Sanjay, advisor
Pasricha, Sudeep, committee member
Malaiya, Yashwant, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Hardware accelerators are optimized functional blocks designed to offload specific tasks from the CPU, speed up them up and reduce their dynamic power consumption. It is important to develop a methodology to efficiently implement critical algorithms on the hardware accelerator and do systematic design space exploration to identify optimal designs. In this thesis, we design, as a case study, a hardware accelerator for the 1-D Finite Difference Time Domain (FDTD) algorithm, a compute intensive technique for modeling electromagnetic behavior. Memory limitations and bandwidth constraints result in long run times on large problems. Hence, an approach which increases the speed of the FDTD method and reduces bandwidth requirement is necessary. To achieve this, we design an FPGA based hardware accelerator. We implement the accelerator based on time-space tiling. In our design, p processing elements (PEs) execute p parallelogram shaped tiles in parallel, each of which constitutes one tile pass. Our design uses a small amount of redundant computation to enable all PEs to start "nearly" concurrently, thereby fully exploiting the available parallelism. A further optimization allows us to reduce the main memory data transfers of this design by a factor of two. These optimizations are integrated in hardware, and implemented in Verilog in Altera's Quartus II, yielding a PE that delivers a throughput of one "iteration (i.e., two results) per cycle". To explore the feasible design space systematically, we formulate an optimization problem with the objective of minimizing the total execution time for given resource constraints. We solve the optimization problem analytically, and therefore have a provably optimal design in the feasible space. We also observe that for different problem sizes reveal that the optimal design may not always match the common sense intuition.
Description
Rights Access
Subject
hardware accelerators
stencil computations