cuda-toy

Check out the write up of this project.

The goal of this project is to implement a simple neural network in C++ that is GPU accelerated via CUDA.

Goals

• [DONE] Implement non-accelerated net
• [DONE] Optimize net architecture a little for MNIST data
• [DONE] Implement CUDA-accelerated matrix functions
• [DONE] Quantitatively compare accelerated vs non-accelerated performance

Results

CPU with std::vector

Mean iteration time ~ 3430 ms

CPU with arrays

Mean iteration time ~ 428 ms

CPU with GPU matrix multiplication

Mean iteration time ~ 18 ms

CPU with GPU matrix arithmetic

Mean iteration time ~ 12 ms