Developer Builds C++/CUDA Deep Learning Framework That Edges Out PyTorch in Benchmark

An AI developer built Aakaar, an open-source deep learning framework written from scratch in native C++ and CUDA with a Python frontend, over the past several months. The framework includes 18 custom loss modules and 11 hand-coded optimizers, with explicit memory contiguity management replacing standard autograd. In a 5-epoch benchmark on the EMNIST dataset using an Intel i7 and RTX 4060 GPU, Aakaar completed training in 127.76 seconds at 83.30% accuracy, compared to PyTorch's 131.23 seconds at 82.55%. The slight speed advantage is attributed to bypassing Python runtime overhead during low-level optimizer operations by keeping them entirely in the C++ backend. The project is publicly available on GitHub, and the developer is seeking feedback from the systems engineering and CUDA optimization communities.