Dr. Jonatan Aronsson discusses how EM solver results are evaluated and how numerical methods affect large 2.5D/3D package simulations.
At IEEE EMC+SIPI 2026 in Dallas, CEMWorks Founder and CEO Dr. Jonatan Aronsson presented two talks examining EM solver performance from different perspectives. One focused on how solver results should be evaluated. The other looked at how numerical methods can affect the practical simulation of modern 2.5D and 3D advanced packages.
Evaluating Solver Performance
In “Evaluating Solver Performance on the IEEE EPS Benchmark Models,” Jon discussed what “solver performance” really means in EM and packaging simulation. His main point was that performance cannot be judged from a single metric.
Matching a measurement does not automatically make a solver “good.” In the same way, completing a simulation faster does not automatically make one solver “better.”
Results depend on the numerical method, the specific benchmark, and the quality of the reference data.
Jon concluded the presentation with a statistical study of a single net in a package. The study showed how manufacturing tolerances, material properties, and probe placement can create variation in the results.
This means that comparisons become more meaningful when this variation is considered, rather than simply looking at whether two individual results match.
Integral-Equation Methods for 2.5D and 3D Packaging
Jon’s second presentation, “Integral-Equation-Based EM Solver for Signal Integrity Analysis in 2.5D/3D Advanced Packaging,” focused on why signal integrity analysis has become increasingly difficult for modern advanced packages.
A single package can span three to four orders of magnitude in feature size within one EM problem. Jon described structures ranging from sub-micron redistribution layers to millimeter-scale board structures and BGA fan-out, together with hundreds of coupled ports, lossy interposers, TSVs, and microbumps.
According to the presentation, this scale can create a major challenge for FEM because its volumetric mesh and surrounding air box grow with the bounding volume. Memory requirements can therefore become a limiting factor before the desired level of accuracy is reached.
Jon then presented an adaptive integral-equation approach that meshes only the surface currents. He concluded with an interposer benchmark comparing the two approaches.
For this specific case, the BEM solver completed the simulation in 11 minutes using 15 GB of memory, compared with 1,265 minutes and 655 GB for FEM. That corresponds to approximately a 115× difference in runtime, while the FEM case required about 44× more memory.
These results represent the specific benchmark presented in Dallas and should not be interpreted as a universal performance ratio between BEM and FEM.
Performance Is More Than One Number
Taken together, the two presentations provide two important perspectives on solver performance. A faster result does not automatically mean a better solver, and a close match with one measurement does not tell the whole story. At the same time, the numerical method can have a major impact on the computational resources required for a particular problem.
Solver performance is therefore not just about speed. It also depends on the numerical method, the benchmark, the reference data, and the variation present in real-world measurements and manufacturing.
These were two of the key themes Jon brought to the discussions in Dallas and important considerations when evaluating EM simulation for increasingly complex electronic packages.