Benchmark Methodology

How we measure and report performance.

Test Environment

Hardware

Component	Specification
GPU	NVIDIA RTX 4090 (24GB VRAM)
CPU	Intel i9-13900K (24 cores)
RAM	64GB DDR5-6000
Storage	NVMe SSD

Software

Component	Version
OS	Ubuntu 22.04 LTS
CUDA	12.4
Driver	550.x
OpenCV	4.8.0
GCC	11.4

Measurement Methodology

Warm-up

Before each measurement:

1. Run operation 10 times to warm up GPU
2. Clear GPU cache with cudaDeviceSynchronize()

Timing

auto start = std::chrono::high_resolution_clock::now();

// Run operation N times
for (int i = 0; i < iterations; i++) {
    operation();
    cudaDeviceSynchronize();
}

auto end = std::chrono::high_resolution_clock::now();
auto avg_time = (end - start) / iterations;

Metrics

• Latency: Average time per operation
• Throughput: Operations per second
• Speedup: CPU time / GPU time

Image Sizes

Name	Dimensions	Pixels
HD	1280×720	921K
FHD	1920×1080	2.1M
4K	3840×2160	8.3M
8K	7680×4320	33.2M

Reproducibility

All benchmarks are available in the benchmarks/ directory:

mkdir build && cd build
cmake -DBUILD_BENCHMARKS=ON ..
make -j$(nproc)
./bin/benchmark_convolution

Notes

• CPU benchmarks use single-threaded OpenCV
• GPU benchmarks include host-device transfer time
• Results may vary based on GPU temperature and clock speed