Femtosecond Floors: Measured Skew Resolution, False Skew and Numerical Jitter of FP32 WebGPU Krylov Transient Simulation on RC and RLC Clock H-Trees, With and Without FP64-Residual Refinement, Checked Against ngspice

WebGPU's shading language (WGSL) has no 64-bit floating-point type, so a browser-based transient simulator must solve the linear system of every time step in single precision (FP32). We measure what this does to clock-tree timing at the femtosecond scale. Perfectly symmetric RC and RLC binary H-trees with 16 and 64 sinks (true skew 0) and copies of the 16-sink tree with known asymmetries of 2.85–2851 fs are integrated with the fixed-step trapezoidal rule (h = 0.25 ps). Each step is solved by a 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. Pure FP32 moves all sinks together (+3.78 fs on the 16-sink RC tree, −204 fs on the 64-sink tree) and reports exactly 0 fs skew on the symmetric trees, a consequence of exact symmetry with our kernel and node ordering rather than a general property. On asymmetric trees it misreports small skews, by +31 % at 2.85 fs and with the wrong sign (−13.4 fs) for a 27.83 fs resistive skew, and at 2 GHz it adds 0.37–19.6 fs RMS of crossing jitter. One warm-started FP64-residual pass at the same CG iteration count per step reproduces all five injected skews within 0.19 % (≤ 0.11 fs), lowers the jitter to ≤ 0.0075 fs RMS, and in a ±1 % wire-resistance Monte Carlo its skew error is 3–4 orders of magnitude below the skew spread (SD 318.6 fs). All WebGPU runs used the SwiftShader CPU software adapter via Node.js and Dawn; no GPU was used and 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.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24357146
Primary Topic
Parallel Computing and Optimization Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Femtosecond Floors: Measured Skew Resolution, False Skew and Numerical Jitter of FP32 WebGPU Krylov Transient Simulation on RC and RLC Clock H-Trees, With and Without FP64-Residual Refinement, Checked Against ngspice

I-Wei Tseng
Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
preprint

Femtosecond Floors: Measured Skew Resolution, False Skew and Numerical Jitter of FP32 WebGPU Krylov Transient Simulation on RC and RLC Clock H-Trees, With and Without FP64-Residual Refinement, Checked Against ngspice

I-Wei Tseng
preprint en

Abstract

WebGPU's shading language (WGSL) has no 64-bit floating-point type, so a browser-based transient simulator must solve the linear system of every time step in single precision (FP32). We measure what this does to clock-tree timing at the femtosecond scale. Perfectly symmetric RC and RLC binary H-trees with 16 and 64 sinks (true skew 0) and copies of the 16-sink tree with known asymmetries of 2.85–2851 fs are integrated with the fixed-step trapezoidal rule (h = 0.25 ps). Each step is solved by a 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. Pure FP32 moves all sinks together (+3.78 fs on the 16-sink RC tree, −204 fs on the 64-sink tree) and reports exactly 0 fs skew on the symmetric trees, a consequence of exact symmetry with our kernel and node ordering rather than a general property. On asymmetric trees it misreports small skews, by +31 % at 2.85 fs and with the wrong sign (−13.4 fs) for a 27.83 fs resistive skew, and at 2 GHz it adds 0.37–19.6 fs RMS of crossing jitter. One warm-started FP64-residual pass at the same CG iteration count per step reproduces all five injected skews within 0.19 % (≤ 0.11 fs), lowers the jitter to ≤ 0.0075 fs RMS, and in a ±1 % wire-resistance Monte Carlo its skew error is 3–4 orders of magnitude below the skew spread (SD 318.6 fs). All WebGPU runs used the SwiftShader CPU software adapter via Node.js and Dawn; no GPU was used and 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.

Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon