Shortly after the launch of the RTX Blackwell generation in Las Vegas, Nvidia revealed that it has a supercomputer entirely dedicated to the continuous improvement of its DLSS (Deep Learning Super Sampling) technology, and has been doing so for six years.
Brian Catanzaro, vice president of applied deep learning research at Nvidia, revealed an impressive infrastructure during a presentation on DLSS 4. “ We have a massive supercomputer at Nvidia, equipped with thousands of our latest GPUs, that runs 24 hours a day, 365 days a year, for the sole purpose of improving DLSS “, he said.
This giant system constantly analyzes DLSS failures in games. When the model fails, producing unwanted effects like ghosting, flickering, or blurring, the team examines these cases in detail to understand the reasons for these errors. These analyzes allow training datasets to be continually enriched with new examples of complex situations that DLSS must handle.
DLSS is constantly evolving thanks to a giant supercomputer
The process is meticulous: after each update of the training data, the model is retrained and then tested on hundreds of different games. This systematic approach explains the considerable progress made since the introduction of DLSS six years ago, culminating in the recent version DLSS 4. The latter abandons convolutional neural networks in favor of a new transformative model.
This revelation comes as Nvidia presented its new RTX 50 Blackwell series a few days ago, which promises high performance thanks to DLSS 4. The RTX 5070, priced at $549, would even achieve performance comparable to that of the RTX 4090 when DLSS is enabled. Of course, when the technology is turned off, the new graphics card falls far short of the performance of the previous generation’s high-end.
