This project is a Fully Connected Neural Network implemented purely in CUDA C.

The neural netrok was trained on BelgiumTS Dataset and used for traffic signs classification.

The purpose of this implementation is to gain a peak performance on a single GeForce GTX 1080 Ti card. It achieves 12.42 times speedup in compartion to 8 core CPU implementation.

• Matrix multiptications, transpositions and reductions performed in shared memory
• Grid-stride loop technique
• Minimization of data transfer between CPU and GPU
• Constant memory utilization
• Streams and asynchronous kernels
• Grid and block sizes optimized per kernel for GTX 1080 Ti
• Pragma unrolling