---
license: mit
tags:
  - gaussian-splatting
  - fault-tolerance
  - single-event-upset
  - reliability
  - radiance-fields
  - computer-graphics
pretty_name: Single-Event Upsets in 3D Gaussian Splatting
configs:
  - config_name: single_bit_upsets
    data_files:
      - split: train
        path: data/single_bit_upsets/*.parquet
  - config_name: multi_upset
    data_files:
      - split: train
        path: data/multi_upset/*.parquet
---

# Single-Event Upsets in 3D Gaussian Splatting Rendering

Artifacts for the paper *Single-Event Upsets in 3D Gaussian Splatting Rendering: Bit-Level Criticality, Spatial Extent, and a Parallel Support Guard* (F. Alpay and B. Basaran).

A trained 3DGS model is a large floating-point array resident in GPU memory, so a single-event upset is one flipped bit in one parameter. This repository releases the fault-injection engine, the trained models, the per-cell aggregated records of more than three million controlled single-bit upsets, the multi-upset records, the throughput measurements, the logs (including the periodic device-utilization trace), and the scripts that regenerate every figure and table in the paper.

## Trained scenes

| scene | primitives | PSNR (dB) | SSIM |
|---|---:|---:|---:|
| chair | 134,826 | 39.22 | 0.9942 |
| lego | 121,691 | 30.70 | 0.9815 |
| ficus | 125,606 | 31.99 | 0.9855 |
| hotdog | 63,452 | 38.50 | 0.9874 |

## Layout

```
code/        training, fault-injection engine, campaign, analysis, figures
models/      trained gsplat checkpoints per scene (model.pt)
data/        per-injection records as Parquet (browsable in the viewer)
results/     aggregate.json, bench.json, multiupset/largescene records, summaries
logs/        campaign / driver / GPU-utilisation logs
figures/     regenerated figures, tables, and numbers.tex
```

Paper: see PAPER_URL below. Reproduce with `code/` following `code/`'s header comments; the GPU run used an RTX 5090 (sm_120), PyTorch 2.12 / CUDA 13, gsplat 1.5.3.

PAPER_URL: __ARXIV_LINK_PLACEHOLDER__

The per-injection records are browsable in the dataset viewer (`single_bit_upsets`: one row per single-bit upset, several million rows; `multi_upset`: accumulated-dose records), stored as Parquet under `data/`.