At CES 2025, Nvidia unveiled its much-anticipated RTX 50-series graphics cards, featuring the revolutionary DLSS 4 technology. We had the opportunity to converse with Bryan Catanzaro, Vice President of Applied Deep Learning Research at Nvidia, who shared exclusive insights into the advanced mechanisms behind DLSS 4, from its innovative transformer model to the newly introduced multi-frame generation (MFG) feature.

Despite it being just over a year since our last discussion — which coincided with the launch of DLSS 3.5 and the acclaimed Cyberpunk 2077: Phantom Liberty — this interview revealed significant advancements that either cater specifically to RTX 50-series users or will be accessible to a broader array of Nvidia graphics cards.

The Shift from CNNs to Transformers: Why Change?

We began by discussing the fundamental transition from convolutional neural networks (CNNs) to transformers. Catanzaro explained, “We’ve been evolving the super-resolution model for about five to six years, and it has become increasingly challenging to pack more intelligence into the same framework. Innovating is crucial, and transformers have demonstrated great scalability, allowing us to train on vast amounts of data. This facilitates smarter model creation and, ultimately, exceptional image quality.

Improving Image Characteristics

When asked about specific improvements in image quality, Catanzaro noted key issues such as stability and ghosting, which plague traditional super-resolution techniques. “Typically, achieving more detail leads to ghosting, while stability issues can cause flickering in distant geometry. The trade-offs we've achieved with our new model greatly enhance this aspect compared to previous iterations, making our graphics significantly more robust.”

Future Potential of Super Resolution

Discussing the trajectory of advancements, Catanzaro asserted, “Larger models trained on high-quality data yield better results. The new transformer model inherently possesses greater capability, allowing for enhanced learning and significantly fewer model failures, such as shimmering and blurring.”

Frame Generation and Latency Reduction

Another fascinating aspect of DLSS 4 is the new frame generation algorithm, which abandons hardware optical flow. “Employing a fully AI-driven solution gives us better results and enhances image quality, particularly for the RTX 50 series,” Catanzaro explained. The changes not only improve performance but also make the algorithm lighter on memory usage.

For gamers, the ultimate user experience has been paramount. With the new frame pacing adaptations, Catanzaro noted an impressive reduction in displayed frame time variability. “We've significantly improved how frames are presented, making gameplay smoother — especially critical when using multi-frame generation.

Reflex 2 Enhancements

Turning our attention to Nvidia's Reflex technology, which now features generative AI capabilities, Catanzaro elaborated, “Reflex 2 dramatically enhances latency by adapting in real-time to camera movements, leading to a more fluid experience and reduced input lag.”

Looking Ahead: The Future of Graphics

Peering into the exciting future of graphics, he suggested that as technology advances towards 1000Hz monitors, AI and neural rendering will play pivotal roles. “We’re just at the starting line of a major evolution in graphics. Traditional bottom-up rendering falls short for complex scenarios; the shift towards neural rendering will empower us to learn from real-world data, making graphics more lifelike and immersive.”

This year marks a turning point for Nvidia and the gaming community. As companies like Nvidia push the boundaries of AI and graphics technology, gamers can expect a future filled with realistic rendering, faster frame generation, and smoother gameplay experiences. With DLSS 4 leading the charge, the world of gaming graphics is set for an exhilarating transformation!