DLSS 4: Frame Multiplication and Revolutionary AI Technology
At CES 2025, NVIDIA took the next step in the evolution of DLSS by introducing DLSS 4. This technology is specifically designed for graphics cards and laptops from the GeForce RTX 50 series, which includes the new Multi Frame Generation feature. From launch, 75 games and applications will support Multi Frame Generation.
Multi Frame Generation generates up to three additional frames for each traditionally rendered frame, working in sync with the full suite of DLSS technologies. This allows for frame multiplication up to 8 times compared to traditional rendering. With GeForce RTX 5090, this technology unlocks stunning performance of 4K at 240 FPS, fully with ray tracing.
In practice, DLSS 4 with Multi Frame Generation doubles performance compared to DLSS 3, as well as reduces latency by half, providing more responsive gameplay and improved image quality.
Revolution in AI Models
DLSS 4 introduces the largest improvement in its AI models since the launch of DLSS 2.0 in 2020. The new components include:
DLSS Ray Reconstruction,
DLSS Super Resolution,
DLAA (Deep Learning Anti-Aliasing),
all powered by transformer models – the same AI architecture that powers leading technologies like ChatGPT, Flux, and Gemini.
The transformer models of DLSS offer better temporal stability, reduced ghosting, higher detail in motion, and smoother edges.
Multi Frame Generation: Technical Innovations
DLSS 3 Frame Generation, used in the RTX 40 series, generated one additional frame by relying on optical flow (Optical Flow Accelerator) and an AI model. However, generating multiple frames was too taxing for the GPU.
DLSS 4 Multi Frame Generation overcomes these limitations by combining innovations in the hardware of the Blackwell architecture and DLSS software. The new AI model is:
- 40% faster,
- uses 30% less VRAM,
- needs to be executed only once for each rendered frame.
For example, in the game Warhammer 40,000: Darktide, the new model provides 10% faster frame rate, using 400 MB less memory at 4K and maximum settings.
5th generation Tensor cores in GeForce RTX 50 GPU provide up to 2.5x higher AI performance, allowing the GPU to process five AI models for each rendered frame within milliseconds.
Hardware and Software Integration
To ensure smoothness, the Blackwell architecture uses Flip Metering – a hardware solution for more precise control of frame display timing. This prevents variability in frame pacing, enhancing the visual experience.
With DLSS 4, NVIDIA creates 15 out of every 16 pixels through AI, achieving an incredible combination of image quality, smoothness, and low latency.
Transformer Model: Improvements in Image Quality
The new transformer model for DLSS replaces the previously used Convolutional Neural Networks (CNN). It:
Uses self-attention operations to analyze each pixel in relation to others within one or more frames.
Has double the parameters of CNN, providing better stability, less ghosting, and higher detail.
Example: In Alan Wake 2, the new model improves the stability of details like metal fences, eliminates ghosting of fans, and reduces flickering of power lines.
For the game Horizon Forbidden West: Complete Edition, the model enhances the detail of textures on the hero's clothing and the overall clarity of scenes.
Compatibility and Availability
DLSS 4 is fully compatible with existing DLSS integrations. When GeForce RTX 50 GPUs hit the market, 75 games and applications will support Multi Frame Generation. Additionally, over 50 games will benefit from the new transformer models.
The games that will support Multi Frame Generation at launch include:
Alan Wake 2,
Cyberpunk 2077,
Star Wars Outlaws,
Black Myth: Wukong, and many others.
Conclusion
NVIDIA DLSS 4 elevates scaling technologies to a new level, providing significant improvements in performance and image quality. Together with GeForce RTX 50 series, this is another major leap in the evolution of gaming and professional applications.
