NVIDIA's NTC Algorithm: Higher-Res Textures, Lower Memory Use

NVIDIA researchers develop new texture compression algorithm with higher resolution and lower memory usage.

NVIDIA researchers have developed a new texture compression algorithm. In GPU graphics rendering, textures play a crucial role. As gaming technology advances, textures become more intricate and larger, resulting in increased consumption of system resources. Therefore, there is a growing need for efficient compression algorithms.

The new algorithm from NVIDIA, called Neural Texture Compression (NTC), utilizes a tensor representation of textures with three dimensions. It compresses multiple channels and Mipmaps together, achieving significantly better compression quality than JPEG XL or AVIF formats.

One notable advantage of the NTC algorithm is that it does not require specific hardware. It leverages matrix multiplication and can be accelerated on any modern GPU (though there may be additional benefits on RTX series GPUs with built-in Tensor cores). Furthermore, it minimizes disk and memory usage.

According to NVIDIA, compared to the traditional GPU texture compression algorithm, BC (Block Compression), the new NTC algorithm brings a fourfold increase in resolution. For example, it allows textures to scale from 1024x1024 to 4096x4096, enlarging the texture scale by 16 times. The difference in quality is substantial.

When compressing a texture with nine channels and a 4K resolution using an RTX 4090, the BC algorithm results in a size of 3.33MB, while the NTC algorithm yields a slightly larger size of 3.6MB. However, the rendering time only increases from 0.49ms to 1.15ms.