Building a Deep Learning Framework in C++ and CUDA: Benchmarking Against PyTorch

Short summary

A developer built Aakaar, a deep learning framework from scratch in C++ and CUDA with a Python wrapper, featuring 18 loss modules and 11 custom optimizers with manual memory management. Benchmarked against PyTorch on EMNIST over 5 epochs, Aakaar achieved 127.76s (83.30% acc) vs PyTorch's 131.23s (82.55% acc), edging out PyTorch by bypassing Python runtime overhead. The framework is open-source and the author seeks feedback from systems engineering and AI infrastructure communities.

• •Built Aakaar, a C++/CUDA deep learning framework with Python wrapper from scratch
• •Benchmarked against PyTorch on EMNIST: 127.76s vs 131.23s with comparable accuracy
• •Key challenge was manual memory contiguity management during backpropagation without autograd

Generated with AI, which can make mistakes.