FP64-Equivalent Transient Waveforms of Lossy RLGC SerDes Channels on FP32-Only WebGPU: Warm-Start FP64-Residual Refinement, Residual Scaling for Multi-Pass Refinement, and Measured Comparison with ngspice
WebGPU's shading language has no 64-bit floating-point type, so a browser-native circuit simulator must solve its linear systems in single precision (FP32). We measure the effect on waveform accuracy in fixed-step trapezoidal transient analysis of a lumped lossy RLGC channel (250 to 2000 segments, up to 4001 unknowns) driven by a 10 Gb/s PRBS7 pattern. Each step is solved by an FP32 Jacobi-preconditioned conjugate-gradient (CG) kernel in WGSL inside FP64 residual refinement on the host. On the 1000-segment channel, the unrefined FP32 solve deviates from an FP64 SciPy reference by up to 1.987 mV, an error that combines the 1e-6 CG tolerance and FP32 rounding. One cold refinement pass lowers it to 1.37e-5 mV; a warm-started single pass reaches 0.0108 mV at 12.0 instead of 10.8 CG iterations per step. Both refined runs differ from ngspice-41 by at most 0.267 mV, the same gap as the FP64 reference. The warm pass matches FP64 waveforms, not FP64 residuals (worst relative residual 8.1e-5). Unscaled repeated refinement failed: the worst residual rose from 5.3e-9 to 7.9e-6 at two passes and became NaN at three, likely through FP32 underflow in CG dot products. Scaling each residual to unit infinity norm restored a monotone decrease to about 1e-13. All WebGPU runs used Chrome's SwiftShader CPU software adapter on a machine without a GPU; timings reflect software emulation, not GPU speed. Drafting was assisted by a large language model; all simulations, 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.24338022
- Primary Topic
- Numerical Methods and Algorithms
- Type
- preprint