Kintech Lab - Intagrated Tools for Inventive Solutions



HiPerCone Company (affiliate of Kintech Lab Ltd.) announces a bunch of new electromagnetic simulation products to be released in 2017-2019. The new software uses revolutionary approach in re-designing traditional Finite Difference Time Domain algorithms to overcome the memory performance bottleneck. This will make large scale full-wave simulations very efficient for any type of modern computational platform ranging from a single CPU or GPU workstation to a supercomputer with tens of thousands of cores. The performance linearly scales with the core number regardless of the memory usage that may exceed the total RAM size of the computer. With the new HiPerCone FDTD software it will be possible to accurately calculate three dimensional full-wave models containing hundreds of wavelengths in each direction. For example, models of human body in microwave range, or of OLED emission layers in optical range, or broadband elipsometry model for nanostructured surface will be calculated within minutes of compute time using typical multi-CPU or multi-GPU workstations.


With the increased parallelism of modern computers reaching millions of cores the Finite Difference Time Domain (FDTD) methods for full wave simulations become very advantageous. Being based on local explicit stencils and equipped with adaptive mesh refinement techniques, they are highly scalable compared to finite elements approaches. However, traditional FDTD mesh update algorithms remain memory bound und thus FDTD codes utilize typically only 5-15% of CPU floating point performance of modern computers. The new Locally Recursive non-Locally Asynchronous (LRnLA) technology re-thinks the data flow in FDTD stencil computations. It accounts for space-time data dependencies determined by characteristics of discrete Maxwell’s equations. Effective usage of computer memory hierarchy is achieved by maximally reusing the loaded mesh data for updates within hyperconic domains of influence. This drastically reduces the number of memory loads/stores per floating point operation.



Resolving the problem of memory restriction


Product Performance

Estimated performance for the new HiPerCone FDTD Solvers is 5-15 times higher than the one achievable in modern traditional CPU- and GPU-based FDTD codes. The exact timings for solving a specific simulation problem depend on the simulation model composition (metals, tensor dielectrics, etc.) and the type of computational hardware used. The estimated times for a typical 3D model problem (10^3 cells in each direction, one third of the cells occupied by a dispersive material, 5x10^3 time steps) are given below for comparison. This typical 3D problem has mesh size large enough for various applications. As examples may serve a reflectometry model for nanostructured surface with dimensions of 10x10x2 microns or corrugated layered structure of emissive layers in LED/OLED with 15x15x1 microns, simulation of full human body model with resolution of 1 mm under the influence of microwave radiation.


Radiation pattern of an OLED with a nanostructured cathode at different wavelengths