No Ramp, but an Undamped Alternating Mode: Measured Long-Horizon FP32 Error of WebGPU Trapezoidal Transient Simulation of a Buck-Converter Output Filter and Power-Delivery Network, With and Without FP64-Residual Refinement, Checked Against ngspice
Year: 2026
Type: preprint
Abstract: WebGPU's shading language has no 64-bit floating-point type, so a WebGPU circuit simulator must solve the linear system of every time step in single precision (FP32). We ask whether this produces a slow drift in long power-integrity transients. A buck-converter output filter, driven by an ideal 2 MHz switch-node source and feeding a board–package–die power-delivery network (14 unknowns; decay-time ratio 208, i.e. moderately stiff), is integrated with the fixed-step trapezoidal rule at h = 2 ns over 100,000 and 500,000 steps. Each step is solved by an FP32 Jacobi-preconditioned conjugate-gradient (CG) kernel in WGSL, alone or inside host-side FP64-residual refinement, and compared with an FP64 SciPy solution on the same grid and with ngspice-41. No configuration shows a linear trend in the cycle-averaged output-voltage error (R² ≤ 0.21 over 0–200 µs, ≤ 0.06 over 0–1 ms). The pure-FP32 error instead sits mainly in an undamped component that changes sign every step: at 990–1000 µs it makes the output ripple read 16.23 mV instead of 5.08 mV, and the cycle-averaged error stays positive (mean +51.3 µV over 200–1000 µs) without ramping. One warm-started refinement pass at almost the same CG cost (14.3–14.8 vs 14.0 iterations per step) limits the maximum output error to 87.5 µV over 1 ms. Its alternating component levels off at 79.5–87.3 µV after 180 µs, and its error against ngspice equals the 0.373 mV SciPy–ngspice difference. Backward Euler suppresses the alternating growth but adds up to 31.4 mV of discretisation error at the die. 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. Part of the KING TSENG AI Web EDA series. Drafting was assisted by a large language model; all simulations (ngspice-41, a SciPy FP64 reference and the WebGPU solver on the SwiftShader CPU software adapter via Node.js and Dawn), numbers and figures were produced and checked by the author's own scripts. No GPU timings are reported.
Author: TSENG, I-WEI
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235999057
