GPU-Accelerated CNN Benchmark

This project explored the practical benefits of GPU acceleration in deep learning. I trained a small CNN on the CIFAR-10 dataset twice: once on CPU and once on an NVIDIA Tesla T4 GPU.

The GPU achieved ~5.5× faster training per epoch and reached 72% accuracy in 20 epochs, compared to the CPU plateauing at ~49% after 2 epochs. Although GPU total runtime was longer, it enabled deeper training in the same wall-clock time.

I implemented the model in PyTorch, measured runtime with synchronized timers, and visualized results with bar charts comparing CPU vs GPU speed and accuracy.

• Implemented a 3-layer CNN in PyTorch for CIFAR-10 classification.
• Benchmarked CPU vs GPU training runtime and accuracy.
• Measured per-epoch speedup (~5.5× faster on GPU).
• Visualized performance with bar charts and accuracy plots.
• Demonstrated why GPU acceleration is critical for modern AI workflows.