Alan Adzima with adviser P. T. Krein

New motor designs are tested with analysis and simulation of their electric fields. Analytic systems used include the finite element method, magnetic equivalent circuits, and boundary element method (BEM). These all require computationally intensive matrix system generation and solving to analyze physical systems. Serial-solver based commercial packages also have lengthy execution times. At the current rate of increasing computer computation processing, it will be decades before real-time simulation can be achieved. Thus new tools are needed.

Parallel processing can distribute a computational load across multiple processing units and dramatically accelerate the process, even on less than cutting-edge hardware. One example is the computer gaming industry’s graphic processing units (GPUs) used in real-time, high-definition three-dimensional graphics. These are highly parallel, hardware multithreaded, many-core processors with tremendous computation power. The challenge has been to unlock this substantial hardware power, because prior GPSs aimed solely at image processing did not support convenient software development in any other context.

With the advent of C-based coding languages for graphic processing units (CUDA — complete unified device architecture–is the best known), existing MATLAB routines can be transitioned to parallel methods by porting functions. The BEM algorithm was chosen to implement CUDA on an electomechanics problem. A laptop with dual core Intel T8300 CPU, NVidia M8600 GT CUDA 1.0 capable GPU, 4 GB of RAM, and running Vista 64bit with MATLAB 64 bit was used. A single-precision N-body simulation to bench mark the system was executed and found to consistently perform at 28 gigaflops for single precision floating-point operations, or a raw computational acceleration of approximately 10x. The CUDA-enabled BEM algorithm provided a 2.5x acceleration factor on sub-matrix generation, leading to an overall performance boost of 1.2x. This modest increase resulted because of the conversion between single-precision and double-precision analysis. Future work includes minimizing data transactions between the device and host in order to maximize overall performance, optimizing algorithms, and searching for methods other than BEM for solving the electromechanical equations.

By adapting current numerical analysis methods to parallel coding techniques that employ modern GPUs, the day of full real-time force and field solutions for electric machines can be moved up ten years or more to help make intensive simulations possible. This report was based on the following publication:

A. J. Adzima, P. T. Krein, T. C. O’Connell, “Investigation of accelerating numerical-field analysis methods for electric machines with the incorporation of graphic-processor based parallel processing techniques,” IEEE Electric Ship Technologies Symposium, 2009, April 2009, pp. 59–64.

Research was supported by the Office of Naval Research grant 490128-933010-191100 and the Grainger Center for Electric Machinery and Electromechanics.