NVIDIA is gaining attention again by releasing concrete performance data for its Neural Rendering technology. Although the technology itself has been introduced before, recent GTC presentations have expanded interest by presenting specific figures on VRAM savings and compression efficiency.
The core technology presented in this presentation is Neural Texture Compression (NTC). Instead of storing texture data as is, it stores it in the form of a small neural network and compressed latent data, and restores it when needed. NVIDIA explained that this can significantly reduce VRAM usage for the same scene, and a demo showed a case where memory usage of about 6.5GB was reduced to less than 1GB. This corresponds to a savings rate of over 80% in simple calculations.
The difference from conventional texture compression methods was also emphasized. Traditional block compression reduces data size but results in quality loss, while NTC compresses texture information in the form of a neural network. It can maintain higher quality for the same capacity and achieve much higher compression ratios for the same quality. This is not just an efficiency improvement, but a change in storage structure.
Another noteworthy point is the nature of the technology. This Neural Rendering series technology is designed to efficiently restore and calculate existing data rather than being based on generative AI. Therefore, it was also emphasized that there is relatively little possibility of issues such as frame distortion or object consistency problems pointed out in recent DLSS 5.
▶ Original source: https://m.kbench.com/?q=node/277602
