📊 Benchmarking the Future: Comparing LabVIEW GPU Toolkits

CuLab, G2CPU, and the Graiphic Accelerator

Welcome to the Graiphic Benchmarking Whitepaper Repository, where we share the methods, results, and LabVIEW sources used to compare the main GPU acceleration toolkits for LabVIEW.

This repository accompanies the official whitepaper:
👉 Benchmarking the Future: Comparing LabVIEW GPU Toolkits CuLab, G2CPU, and the Graiphic Accelerator (v1.1)

---

Overview

This benchmark measures and compares the performance, integration, and determinism of several LabVIEW GPU toolkits — all tested in the same LabVIEW environment.

Toolkits Compared

• Graiphic Accelerator Toolkit
• CuLab GPU Toolkit 4.1.2.80 (Ngene)
• G2CPU GPU and CPU HPC Toolkit 1.6.0.15 (Natan Biesmans)
• Native LabVIEW CPU execution

The objective is to provide a real-world comparison and understand the trade-offs between speed, scalability, and ease of integration.

---

Test Environment

Component	Specification
OS	Windows 11
CPU	Intel® Core™ i9-10850K @ 3.60 GHz
GPU	NVIDIA GeForce RTX 3060
LabVIEW	2025 Q3
CUDA	12.8
TensorRT	10.13.3.9
DirectML	1.15.4.0
Date	November 6, 2025

This setup represents a balanced workstation configuration for reproducible LabVIEW GPU benchmarks.

---

Benchmarks Included

1. GEMM Processing
   Matrix multiplication followed by arithmetic post-processing.
2. Arithmetic Operations
   Iterative Add / Neg / Mul / Div loops for element-wise operations.
3. Complex Number Computation
   Handling of real + imaginary tensors using ONNX custom nodes.
4. Signal Processing Application
   FFT + arithmetic operations on real NI-like signal data (~32 k samples).
   ➤ This test was designed to reflect realistic, small-scale sensor workloads — not synthetic stress tests.

---

Key Findings

• Graiphic Accelerator (TensorRT) achieves the highest performance, up to:

  • 5× faster than CuLab
  • 40× faster than G2CPU

• Compiled-graph execution (ONNX Runtime) drastically reduces overhead compared to per-node DLL execution.

• Complex-number support works using custom ONNX nodes — an area for future standardization.

• For small data blocks, CPU execution remains competitive; GPU benefits increase with workload size.

---

Source Files

All LabVIEW VIs used to generate the benchmark results are available in the
/Source directory.

Benchmark	Folder	Description
GEMM	Source/GEMM	Matrix-multiplication tests
Arithmetic	Source/Not Complex	Scalar & vector operations
Complex	Source/Complex	Custom complex-number computation
Signal Processing	Source/Signal Processing Without Indicator And Warmup	FFT-based signal test

Additional required file:

• TEMP.BIN (2 GB, test data)
  👉 http://download2.graiphic.io/_Bench/TEMP.BIN

---

Replication & Discussion

This benchmark was built for transparency, reproducibility, and collaboration.

Community contributions encouraged:

• Independent replication
• Comparative pull requests
• New test proposals
• Methodology discussions

Discussion board:
https://github.com/Graiphic/whitepapers/issues

Repository:
https://github.com/Graiphic/whitepapers

---

About Graiphic

Graiphic develops the first ecosystem unifying AI + Logic + Hardware + Energy inside a single ONNX graph.

Contact

• [email protected]
• https://www.graiphic.io

---

Versioning

Version	Date	Author	Description
1.0	2025-10-15	Youssef Menjour	First release
1.1	2025-11-07	Youssef Menjour	Added DirectML EP

---

Towards a Community Standard: LabVIEW Open Benchmark Suite (LOBS)

Following this benchmark, we launched
👉 LabVIEW Open Benchmark Suite (LOBS)

LOBS provides:

• Open-source vendor-neutral tests
• Reproducible pipelines
• Transparent comparison criteria

This whitepaper is Reference 0 of the suite.