The Electrical Interconnect and Packaging (EIP) group provides best-of-breed electrical designs, modeling and simulation tools,as well as characterization techniques for new products and future packaging needs in supercomputers and servers. One of the recent additions to this suite of electromagnetic (EM) modeling tools is a full-wave, arbitrary shape solver named EMSurf, which stands for ElectroMagnetic Surface modeling. Through the state-of-the-art numerical techniques, researchers can perform a larger and more complex EM analysis and more accurately analyze electrical performance for improved estimates of noise sources and improved component models.
Modeling electromagnetic performance requires solving Maxwell’s differential equations, which relate current, charge, electric and magnetic fields. Solving these numerically means constructing and solving a set of linear algebraic equations. The two approaches to this are a volumetric partitioning of the geometry, which produces a large number of unknowns but a very sparse linear system, or a surface partitioning, which only partitions the surfaces of the objects and produces significantly fewer unknowns but a dense system matrix. A modification to the surface approach, called pre-corrected FFT, uses surface partitioning to keep the number of unknowns small, but avoids the storage of a dense matrix by projecting all unknown surface currents onto a Cartesian grid of point sources. Multiplication of a vector times this linear system matrix becomes a discrete convolution, and can be done with Fast Fourier Transforms (FFT) requiring that only O(N) matrix elements be stored. The storage requirements become proportional to that of a sparse volumetric technique, with the number of unknowns being that of a surface technique.
The key to the performance of these methods is that the linear system be solved iteratively. One cannot compute the inverse of a matrix that was not stored. The hurdle is that the condition number of the system matrix is extremely poor for geometries where the feature sizes are orders of magnitude smaller then the electromagnetic wavelength. This is exactly the case for all interconnect and packaging problems. The key to achieving convergence, and obtaining a solution at all, is a robust preconditioner for the iterative linear system solver. The EIP group has patented a new preconditioning technique for this project that is the cornerstone of its success. The new preconditioner sections the geometry, solving Maxwell’s equations within each section separately, but enforces the currents at the boundaries of adjacent section to be continuous.
IBM pioneered the first electromagnetic modeling tools in the early 1970's and since it has built up an extensive suite of extremely powerful and accurate solvers. These tools have been verified for accuracy with thousands of practical cases, and are continually evolving to encompass new techniques. One such new technique is the pre-corrected FFT method used in EMSurf, which now allows us to model the largest problem sizes in the industry. For typical problem sizes, EMSurf has demonstrated a speed up of 30x and a memory usage reduction of 10x. For larger problems, the benefit increases due to the nearly linear scaling of the solve time and memory requirements. Also demonstrated was the ability to solve one million surface unknowns, an impossible task for the same class of conventional solvers.
Related Publications
J. Morsey. A broadband, low storage preconditioning scheme based on reduced coupling for full-wave method of moments solvers. Applied Computational Electromagnetics Society. April 2004.
Innovator's corner
Jason Morsey Researcher
What is the most exciting potential future use?
I believe the most exciting potential use is in the parallelization of the tool. Once parallelized, we have the potential to solve truly huge problems using supercomputers such as BlueGene.
What is the most interesting part of your research?
Interacting with the internal users of EMSurf and using it to model the structures that they bring forward.
What inspired you to go into this field?
I would have to give credit to a professor I had during my undergrad and master degree. He was such a good professor that he inspired me to begin my research in computation EM. I suppose fate took over from there.
What is your favorite invention of all time?
The Internet.
Research team
Related Research
Disciplines: Electrical Engineering
Research Areas: Electrical Interconnect and Packaging
Research Labs: Watson Research Center