Neural Networks in High Performance Graphics

Thomas Müller


Please LOG IN to view the video.

Date: November 30, 2022


Neural networks have a reputation for being expensive to run and even more expensive to train, which makes them seem like a bad fit for high-performance tasks. But this is a misconception. Neural networks can run, and even be trained, in the inner loops of real-time renderers or SLAM systems when powered by the right data structures and algorithms. This talk is about these algorithms: why do they work? When does it makes sense to use them? And how important is such low-level engineering in a research project?

Further Information:

Thomas is a principal research scientist at NVIDIA working on the intersection of machine learning and light transport simulation. His research won multiple best paper awards and is used in movie production (such as in Disney’s Hyperion and RenderMan), 3D reconstruction and gaming. As part of his research, Thomas created several widely used open source frameworks, including instant-ngp, tiny-cuda-nn, and tev. Thomas holds a PhD from ETH Zürich & Disney Research and, in a past life, also developed large components of the online rhythm game “osu!”.


Created: Wednesday, November 30th, 2022