BenchLab

BenchLab is a modular GPU benchmarking suite written in C++ using both CUDA and HIP backends. It provides comparative performance benchmarks for common GPU operations across NVIDIA and AMD GPUs.

Features

• Unified interface for HIP (AMD) and CUDA (NVIDIA)
• Modular benchmarks:
  • Vector Addition
  • Matrix Multiplication
  • 1D Convolution
  • Reduction (Sum)
  • Memory Bandwidth Test

---

Build Instructions

Prerequisites

• CMake ≥ 3.10
• A supported compiler (GCC, Clang, or MSVC)
• For CUDA backend:
  • NVIDIA GPU with CUDA Toolkit installed
• For HIP backend:
  • AMD GPU with ROCm installed

Build

Use the provided build.sh script:

# Build for CUDA
./build.sh -f --backend=CUDA

# Build for HIP
./build.sh -f --backend=HIP

# If you don't provide the backend, it will choose the backend based on installed drivers
./build.sh -f

Or build manually with Cmake:

# CUDA
cmake -DENABLE_CUDA=ON -DENABLE_HIP=OFF -B build
cmake --build build -j

# HIP
cmake -DENABLE_CUDA=OFF -DENABLE_HIP=ON -B build
cmake --build build -j

Run

./build/run_benchmarks