Sub-LSB Comparator Decisions on a Real Cloud GPU: Measured FP32 Error of WebGPU Transient Simulation of SAR-ADC Capacitive-DAC Settling on an NVIDIA Tesla T4, With FP64-Residual Refinement and Residual Scaling, Checked Against ngspice

Successive-approximation-register (SAR) analog-to-digital converters decide each bit by comparing the settled top plate of a capacitive DAC (CDAC) with a common-mode voltage, often with margins far below one LSB. The WebGPU shading language has no 64-bit floating-point type, so a WebGPU transient solver must perform its linear solves in FP32. We measure whether such a solver preserves SAR comparator decisions on real GPU hardware. Two generic binary CDACs (12 and 16 bits, 32 copies each, including near-threshold inputs; 1792 decisions per configuration over four cases) are integrated with the trapezoidal rule at h = 1 ps. Each step is solved by a Jacobi-preconditioned conjugate-gradient kernel in WGSL, either alone or inside host-side FP64-residual refinement with or without residual scaling. All 20 headline runs used an NVIDIA Tesla T4 on Google Colab (driver 580.82.07, Node.js 22.23.4 with Dawn on Vulkan, adapter isFallbackAdapter=false). Decisions are compared with a same-grid SciPy FP64 solution that agrees with ngspice-41 to within 1.0e-6 µV. Pure FP32 is off by 81.5–391 µV (0.74–16.5 LSB) and flips 10–67 decisions per case. One scaled FP64-residual pass after a warm start lowers the error to at most 0.145 µV or 0.0034 LSB, leaving a single flip at a 0.0188 µV margin; a cold start with two scaled passes reaches 2.4e-6 µV with no flips. Without scaling, warm refinement produced non-finite values in 3 of 4 cases on the T4, but in only 1 of 4 under SwiftShader CPU emulation. At these sizes (n ≤ 576) the T4 was 2.3–12.5 times slower per time step than SciPy FP64 sparse LU on the same Colab CPU, so no speed advantage is 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.24360042
Primary Topic
Numerical Methods and Algorithms
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Sub-LSB Comparator Decisions on a Real Cloud GPU: Measured FP32 Error of WebGPU Transient Simulation of SAR-ADC Capacitive-DAC Settling on an NVIDIA Tesla T4, With FP64-Residual Refinement and Residual Scaling, Checked Against ngspice

I-Wei Tseng
Zenodo (CERN European Organization for Nuclear Research)
Numerical Methods and Algorithms
preprint

Sub-LSB Comparator Decisions on a Real Cloud GPU: Measured FP32 Error of WebGPU Transient Simulation of SAR-ADC Capacitive-DAC Settling on an NVIDIA Tesla T4, With FP64-Residual Refinement and Residual Scaling, Checked Against ngspice

I-Wei Tseng
preprint en

Abstract

Successive-approximation-register (SAR) analog-to-digital converters decide each bit by comparing the settled top plate of a capacitive DAC (CDAC) with a common-mode voltage, often with margins far below one LSB. The WebGPU shading language has no 64-bit floating-point type, so a WebGPU transient solver must perform its linear solves in FP32. We measure whether such a solver preserves SAR comparator decisions on real GPU hardware. Two generic binary CDACs (12 and 16 bits, 32 copies each, including near-threshold inputs; 1792 decisions per configuration over four cases) are integrated with the trapezoidal rule at h = 1 ps. Each step is solved by a Jacobi-preconditioned conjugate-gradient kernel in WGSL, either alone or inside host-side FP64-residual refinement with or without residual scaling. All 20 headline runs used an NVIDIA Tesla T4 on Google Colab (driver 580.82.07, Node.js 22.23.4 with Dawn on Vulkan, adapter isFallbackAdapter=false). Decisions are compared with a same-grid SciPy FP64 solution that agrees with ngspice-41 to within 1.0e-6 µV. Pure FP32 is off by 81.5–391 µV (0.74–16.5 LSB) and flips 10–67 decisions per case. One scaled FP64-residual pass after a warm start lowers the error to at most 0.145 µV or 0.0034 LSB, leaving a single flip at a 0.0188 µV margin; a cold start with two scaled passes reaches 2.4e-6 µV with no flips. Without scaling, warm refinement produced non-finite values in 3 of 4 cases on the T4, but in only 1 of 4 under SwiftShader CPU emulation. At these sizes (n ≤ 576) the T4 was 2.3–12.5 times slower per time step than SciPy FP64 sparse LU on the same Colab CPU, so no speed advantage is 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)
Numerical Methods and Algorithms
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