A Performance Model for GPU-Accelerated FDTD Applications

Multiple-Gpu-Based Frequency-Dependent Finite-Difference Time Domain Formulation Using Matlab Parallel Computing Toolbox

“…The D fields do not exchange between GPUs. This data-exchange approach was applied in [3] and [8], and in the multi-CPU-core FDTD [16]. In MATLAB, this exchange can be defined by a data movement direction within the first SPMD statement.…”

“…There are two major approaches to programming multi-GPU based parallel FD-FDTD: open computing language (OpenCL) [3,4] and compute unified device architecture (CUDA) [5][6][7][8]. OpenCL is a framework and programming language that executes across heterogeneous platforms consisting of CPUs, GPUs, or other processors.…”

“…Different types of GPUs were used to compare their real computability. Researchers from Belgium used CUDA on up to four GPUs for a 3-D FDTD application with multi-pole dispersion for plasma [7,8], and found an almost linear speedup with respect to the number of GPUs.…”

A Python Framework for Fast Modelling and Simulation of Cellular Nonlinear Networks and other Finite-difference Time-domain Systems

“…For such networks, a viable theory for locating emergent behaviors in the parameter space (or gene's space) called local activity theory [5] was proposed and successfully tested [4]. Fluid dynamics, sound propagation, and many other physical phenomena can be modeled in FDTD frameworks such as cellular automata and Lattice Boltzmann Machines [6][7] [8]. Such models need convenient informatic implementations (modelling and simulation frameworks = MSF), and in recent years various commercial or noncommercial solutions were offered, most struggling to offer GPU support and high performance (short simulation times for wide arrays of cells).…”

A GPU performance estimation model based on micro-benchmarks and black-box kernel profiling

“…However, this inspection suggests that kernels are bound on the DRAM latency as the requested transactions cannot fully utilize the DRAM resource. In contrast, the Tesla GPUs exhibited higher rates on DRAM utilization due to the ECC protection, which caused a much larger DRAM traffic (3,840,153,10 7) alu_fu ( 4) l2 ( 10) dram ( 8) l2 ( 8) dram ( 6) btr-fnd alu_fu (3) alu_fu ( 3) l2 ( 5) single_precision_fu ( 3) alu_fu ( 4) ldst_fu (4) btr-rng alu_fu ( 4) ldst_fu ( 4) l2 ( 6) tex ( 4) alu_fu ( 4) ldst_fu ( 5) bp-adj l2 (4) dram ( 5) l2 ( 9) dram ( 8) dram ( 5) dram (4) bp-fwd alu_fu ( 7) alu_fu ( 7) single_precision_fu ( 6) single_precision_fu ( 5) alu_fu ( 8) alu_fu ( 7) bfs-k1 alu_fu (2) ldst_fu ( 2) dram ( 3) dram ( 2) dram ( 4) dram (3) bfs-k2 alu_fu ( 4) alu_fu ( 3) dram ( 4) dram ( 3) alu_fu ( 4) alu_fu (3) dwt-cpy dram ( 9) dram ( 8) dram (…”