Neural Radiance Fields (NeRF) have emerged as a groundbreaking technique for representing and rendering 3D scenes, allowing for creating high-quality new views from a set of 2D images. In this study, we present a novel approach to speed up NeRF’s positional encoding by utilizing the memory hierarchy of GPUs. Our method utilizes vector quantization to compress positional encoding and stores the resulting codebook in Static Random Access Memory (SRAM), providing much higher access speeds compared to High Bandwidth Memory (HBM). This leads to a concise representation where each feature vector is indexed, greatly reducing memory usage. During inference, only the codebook index, rather than the entire feature vector, is retrieved from HBM, resulting in a 10-fold acceleration in positional encoding. To further enhance rendering quality, we introduce a weighted sum of the codebook vectors to approximate the original feature vector, which allows for a balance between rendering quality and inference speed.