CUDA 13 Fraud AI Benchmark: CPU vs CUDA 12 vs CUDA 13 model-quality comparison

Accelerated Computing CUDA CUDA Programming and Performance
, , ,
1

Hi NVIDIA Developer Community,

I wanted to share an open-source CUDA / PyTorch benchmark repo focused on fraud AI and graph-AI workloads against public fraud datasets:

The project is designed to compare the same fraud-detection models across three environments:

  1. CPU baseline

  2. Older CUDA GPU environment using CUDA 12.6.3

  3. Latest CUDA GPU environment aligned with CUDA 13.3 / CUDA 13.3 Update 1

The benchmark is intended to run against public datasets such as:

  • Credit Card Fraud Detection dataset

  • Elliptic Bitcoin transaction graph dataset

The benchmark intentionally focuses only on metrics that are comparable across all three environments:

  • Accuracy

  • Precision

  • Recall

  • F1 score

  • Parameter count

  • Model size in MB

It deliberately avoids speed, throughput, CUDA kernel timing, GPU utilization, and peak CUDA memory in the primary benchmark because those are not equally comparable across CPU, older CUDA, and latest CUDA environments.

The repo includes:

  • CUDA 13 Dockerfile

  • CUDA 12 Dockerfile for older-version comparison

  • CPU / CUDA 12 / CUDA 13 benchmark matrix

  • --output option to save measured benchmark results

  • Public fraud dataset examples

  • RAPIDS/cuGraph and PyTorch Geometric examples

  • Self-hosted GPU runner guide

  • Kubernetes deployment reference

I would appreciate feedback from the CUDA community on:

  1. Whether the CUDA 12 vs CUDA 13 comparison setup is technically sound.

  2. Whether the benchmark design against public fraud datasets is appropriate.

  3. Whether the selected metrics are suitable for model-quality comparison.

  4. Whether the Docker setup is aligned with good CUDA reproducibility practice.

  5. Any suggestions to improve the repo before adding measured CUDA 12 and CUDA 13 benchmark outputs.

The current goal is not to claim performance improvement, but to create a reproducible CUDA 13-ready fraud-AI benchmark structure against public fraud datasets where measured outputs can be added after running on real GPU environments.

Thanks for any feedback or suggestions.

4

Can you tell which Cuda 13 features you have been using in the newer version?

5

Thanks for asking. For our large-scale graph workload of approximately 700 million nodes and 20 billion edges, we are evaluating GPU-to-GPU communication using the newer CUDA-compatible RAPIDS stack.

cuGraph uses RAFT communicators with NCCL and UCX to exchange data when connected vertices are distributed across different GPUs. Newer versions of this communication stack may improve efficiency, but the actual benefit depends on GPU topology, NVLink/NVSwitch or InfiniBand availability, communication patterns, and graph-partition quality..

6

I get the feeling that the complete project is AI generated.

Has the project been run in any actual real-life installations?