tiny-cuda-nn

Why we included this project

tiny-cuda-nn is the framework behind NVIDIA's Instant NGP, and that alone tells you what it is for: a compact C++/CUDA library that trains and queries small neural networks fast enough for real-time work. Its two signature pieces are a fully fused multilayer perceptron and a multiresolution hash encoding, each covered by its own technical paper, and together they let a network run as a single fused kernel instead of stepping back into TensorFlow or PyTorch layer by layer. The README benchmarks make the payoff concrete, measuring against TensorFlow 2.5 with XLA on an RTX 3090, and the gap is large. Model setup happens through JSON, so swapping an encoding, loss, or optimizer does not require touching code, and PyTorch bindings keep it usable from Python. Rendering and 3D reconstruction are the obvious home, where a trained model gets queried millions of times per frame, but the same fused-kernel speedup applies anywhere you push small, dense networks hard.