Arithmetic-Limited Stopband Dynamic Range of FP32 WebGPU Transient Simulation: Measured Noise Floors of Doubly Terminated LC Ladder Filters With and Without FP64-Residual Refinement, Checked Against ngspice AC Analysis
Year: 2026
Type: preprint
Abstract: WebGPU's shading language has no 64-bit floating-point type, so a WebGPU circuit simulator must solve its linear systems in single precision (FP32). We measure what this costs in stopband dynamic range. Doubly terminated LC ladder low-pass filters (9th- and 15th-order Chebyshev, 1 GHz cut-off) are driven by a coherently sampled two-tone signal at 0.2 and 3 GHz. A fixed-step trapezoidal transient solves each step with an FP32 Jacobi-preconditioned conjugate-gradient (CG) kernel in WGSL, optionally inside FP64-residual refinement, and the output spectra are checked against ngspice-41 AC analysis and an FP64 SciPy transient. Without refinement the median floor is −165 to −168 dB per 10 MHz bin, with band-edge spurs up to −102.6 dB. Tightening the CG tolerance from 1e-6 to 1e-8 changes the floor by at most 2.4 dB, so it is set by FP32 arithmetic. The 15th-order filter's 3 GHz tone (−213.3 dB by AC analysis) is therefore not resolved. One warm-started refinement pass, at the same number of FP32 solves per step, lowers the floor to about −222 dB/bin but reads that tone 18.5 dB high; one cold pass or two warm passes reproduce the FP64 reference measurement (floors −262.6 and −244.8 dB/bin, tone −213.6 dB). All WebGPU runs used the SwiftShader CPU software adapter via Node.js and Dawn on a machine without a GPU; no speed results are claimed. Drafting was assisted by a large language model; all simulations, numbers and figures were produced and checked by the author's own scripts.
Author: TSENG, I-WEI
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236376036
