The paper introduces a versatile and efficient input encoding method for neural graphics primitives, which significantly reduces the cost of training and evaluation. The encoding, called multiresolution hash encoding, is a cascade of grids that map to fixed-size arrays of feature vectors, optimized through stochastic gradient descent. This method allows for the use of smaller neural networks without sacrificing quality, reducing floating-point and memory operations. The multiresolution structure helps disambiguate hash collisions, making the architecture simple and easy to parallelize on modern GPUs. The authors implement the system using fully-fused CUDA kernels to minimize bandwidth and compute operations, achieving a speedup of several orders of magnitude. The encoding is task-agnostic, with the same implementation and hyperparameters used across various tasks, including gigapixel image representation, signed distance functions, neural radiance caching, and neural radiance and density fields (NeRF). The method demonstrates rapid training, high quality, and simplicity, with clear benefits in terms of speed and performance.The paper introduces a versatile and efficient input encoding method for neural graphics primitives, which significantly reduces the cost of training and evaluation. The encoding, called multiresolution hash encoding, is a cascade of grids that map to fixed-size arrays of feature vectors, optimized through stochastic gradient descent. This method allows for the use of smaller neural networks without sacrificing quality, reducing floating-point and memory operations. The multiresolution structure helps disambiguate hash collisions, making the architecture simple and easy to parallelize on modern GPUs. The authors implement the system using fully-fused CUDA kernels to minimize bandwidth and compute operations, achieving a speedup of several orders of magnitude. The encoding is task-agnostic, with the same implementation and hyperparameters used across various tasks, including gigapixel image representation, signed distance functions, neural radiance caching, and neural radiance and density fields (NeRF). The method demonstrates rapid training, high quality, and simplicity, with clear benefits in terms of speed and performance.