README.md

C++ MNIST Benchmark

What It Does

This program uses LibTorch (the C++ version of PyTorch) to train a simple neural network on the MNIST dataset. It performs the following steps:

1. Downloads the MNIST dataset.
2. Loads the entire dataset into GPU memory.
3. Trains a model for 10 epochs.
4. Collects and saves performance stats (time, accuracy, GPU usage, memory) to the stats/ directory in the project root.

How to Run

Prerequisites

• An NVIDIA GPU with CUDA.
• A C++ compiler, cmake, and libcurl.
• A downloaded copy of LibTorch. You must set the LIBTORCH environment variable to point to the extracted LibTorch directory.

Run Commands

From the cpp/ directory, run:

# Set this to your LibTorch path
export LIBTORCH=/path/to/your/libtorch

# Build the project
cd cpp_benchmark
mkdir -p build
cd build
cmake ..
make

# Run the benchmark
./mnist_benchmark_cpp