GraphRTX

GraphRTX is a high-performance RT-accelerated graph analytics framework built on CUDA and NVIDIA OptiX. It accelerates traditional graph algorithms by mapping their core traversal and intersection patterns onto ray tracing hardware (RT cores).

• C++ core engine with direct CUDA/OptiX integration
• Python bindings that load .mtx files or existing networkx graphs
• Out-of-core memory manager for analyzing graphs larger than GPU memory

---

Overview

GraphRTX loads standard graph datasets (.mtx Matrix Market files), builds GPU-optimized adjacency structures (UASP, triangles, and AABBs), and runs a suite of graph algorithms via either:

• a pure CUDA pipeline, or
• a hybrid OptiX + CUDA pipeline that offloads neighbor traversal and intersection work to RT cores.

---

Supported Algorithms

Algorithm	CUDA	Hybrid (OptiX + CUDA)	Notes
BFS — Breadth-First Search	✓	✓	Parallel frontier expansion
PR — PageRank	✓	✓	Configurable damping factor
SSSP — Single-Source Shortest Path	✓	✓	Weighted relaxation
BC — Betweenness Centrality	✓	✓	Approximate version supported
TC — Triangle Counting	✓	✓	Geometric intersection variant
CDLP — Community Detection (Label Propagation)	✓	✓	RT-accelerated neighbor scans
WCC — Weakly Connected Components	✓	✓	RT-accelerated neighbor scans
LCC — Local Clustering Coefficient	✓	✓	Built on TC

Hybrid variants split work between CUDA and OptiX to accelerate neighbor traversal and intersection.

---

Python Example (NetworkX)

import networkx as nx
from pygraph_rtx import Graph

G = nx.fast_gnp_random_graph(1000, 0.01)

grtx = Graph()
grtx.from_networkx(G)
grtx.prepare()

print(grtx.run_bfs(0))

---

Build

Prerequisites

• CUDA ≥ 13.0
• OptiX SDK ≥ 9.0
• CMake ≥ 3.13
• GPU with RT cores (e.g. RTX 4070)

Tested on RTX 4070, RTX 5090, and RTX A6000 Pro.

Compile

Set OPTIX_ROOT to your OptiX SDK install path, then:

git clone https://github.com/anongraph/graphrtx.git
cd graphrtx
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j

See bench/README.md for benchmark CLI options.

---

Evaluation Datasets

General: roadUSA, livegraph, kron21, weibo, twitter10, orkut

LDBC: datagen, graph500, sk-2005, friendster