Resonance-Aware Shifted PRIMA Macromodels for Browser-Native Transient Analysis of On-Chip RLC Power-Delivery Networks: Measured Accuracy Against ngspice and FP32-Safe Reduced Stepping
Browser-native circuit simulation is limited by tab memory and by WebGPU's single-precision (FP32) arithmetic, which makes repeated full-order transient analysis of large on-chip power-delivery networks (PDNs) unattractive. We study passive Krylov model-order reduction (PRIMA) as a compact macromodel for this setting and make two practical changes. First, the block Arnoldi expansion point is placed on the real axis at half the package-die LC resonance (s0 = omega_LC/2), estimated in O(n) from the netlist. Second, the reduced trapezoidal propagator is formed in FP64 and only then cast to FP32 for stepping. On RLC power meshes with 25 pads and 8 pulsed loads (425 to 40,025 unknowns), a 40-state shifted model stays below 0.3% maximum transient error up to 10,025 unknowns, whereas DC-expanded PRIMA needs 48 states for the same level. Against ngspice-41 the 40-state model deviates by at most 0.146 mV (up to 19,906 unknowns). FP64-formed/FP32-stepped propagation lowers the FP32 error 237-669x relative to a naive FP32 pipeline. The reduced model occupies 8.96 KB versus 8.0-43.5 MB for the full sparse LU factors, and a 50-pattern what-if sweep on a 100x100 mesh runs 212x faster than a full-order solver that reuses its LU factors. All results come from a CPU Python/NumPy/SciPy prototype; no browser, WebAssembly or WebGPU build was run. Part of the KING TSENG AI Web EDA series. Drafting was assisted by a large language model; all simulations (ngspice-41 and a Python/NumPy/SciPy prototype), numbers and figures were produced and checked by the author's own scripts.
Authors
- I-Wei Tseng
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24331882
- Primary Topic
- VLSI and FPGA Design Techniques
- Type
- preprint