Applying AI to Ray Tracing for Rendering Prediction
Demonstrating High-Quality Denoising Work with Quadro GPU

NVIDIA announced that it is accelerating graphics work speed by applying artificial intelligence to Ray Tracing.

Jensen Huang, NVIDIA's founder and CEO, demonstrated during the keynote speech at NVIDIA's GPU Technology Conference how applying artificial intelligence to ray tracing advances iterative design processes and enables accurate prediction of final rendering. Ray Tracing is a technology that uses complex mathematics to realistically simulate how lighting interacts with surfaces in a specific space.

While the ray tracing process generates highly realistic images, it is a computationally intensive task that can leave slight noise in the image. The process of removing this noise while maintaining fine edge processing and texture is called Denoising in the industry.

CEO Jensen Huang demonstrated using NVIDIA Iray how NVIDIA combines deep learning prediction algorithms with NVIDIA Quadro GPU based on Pascal architecture to process high-quality denoising work in real time.

This is expected to have a significant impact on graphics-intensive industries such as entertainment, product design, manufacturing, architecture, and engineering. This technology can be applied to various types of ray tracing systems. NVIDIA is already in the process of integrating deep learning technology into its own rendering products, with Iray being the starting point.

To this end, NVIDIA's research and engineering teams focused on a neural network called an autoencoder. Autoencoders are used for image resolution enhancement, video compression, and various image processing algorithms.

NVIDIA's research team trained the neural network to interpret noisy images as clean reference images using the NVIDIA DGX-1 AI supercomputer. They used 15,000 pairs of images with varying levels of noise across 3,000 different scenes for neural network training, which took less than 24 hours. The trained neural network removed noise from nearly all images in less than one second, and the same applied to images not included in the training set.

The Iray deep learning function will be included in the Iray SDK that NVIDIA provides to software companies and will be unveiled in the Iray plugin product scheduled for release later this year. NVIDIA also plans to add an artificial intelligence mode to NVIDIA Mental Ray. Details on this technology will be unveiled at the ACM SIGGRAPH 2017 computer graphics conference to be held in July.