NVIDIA’s DLSS 4 “Multi Frame Generation” technology boosts performance by up to 8x in the most demanding games featuring path tracing. While the finer details are still under wraps, here’s a breakdown of what enables this feat and how it differs from DLSS 3. In a nutshell, NVIDIA has replaced most of its neural networks with updated or revamped AI models that improve the quality and performance of the algorithm.
Updated Frame Generation AI Models
DLSS 3 Frame Generation leverages engine data, including motion vectors, depth, jitter offsets, and the optical flow field from the Optical Flow Accelerator to generate an additional frame between every two frames. This is called frame interpolation. Running this process multiple times for more than one interpolated frame is unfeasible as it would slow down the GPU.
DLSS 4 frame generation mode is 40% faster and uses 30% less VRAM. Furthermore, it is can generate up to 3 frames in one go without calling on the optical flow accelerator. The optical flow field is generated by an AI model, allowing the insertion of multiple frames every two rendered frames.
Hardware Flip Metering & 5th Gen Tensor Cores
Hardware Flip Metering: Frame generation on the RTX 40 series GPUs used CPU-level pacing which can lead to inconsistent frame pacing between frames, resulting in uneven pacing in some cases. This can adversely affect smoothness even if you’re getting high framerates.
The Blackwell GPUs rely on hardware Flip Metering which offloads the queued frame processing to the GPU display engine. This allows the GPU to adjust the frame display timing with increased precision. To facilitate this at optimal performance, Blackwell’s display engine is twice as fast as Ada to support higher resolutions and refresh rates for hardware Flip Metering with DLSS 4: Up to 480Hz at 4K or 120Hz at 8K.
5th Gen Tensor Cores: DLSS 4 employs 5 AI models, including upscaling, ray reconstruction, and multi frame generation between every two rendered frames. The RTX 50 family leverages the 5th Gen Tensor cores which offer 2.5x more throughput than their predecessors. This is achieved by switching to a lower precision FP4 metric.
DLSS Transformer Model
DLSS 4 replaces the widely used CNN models with the transformer model. This vision transformer features “self-attention” operations that evaluate the relative importance of different pixels across a frame and over multiple frames. Using twice as much data as CNN allows transformer models to improve temporal stability, reduce ghosting, and increase the detail perceived in motion.
The newer transformer model is compatible with games featuring the CNN model and will be available by the month’s end. Titles that haven’t been updated to the latest model can utilize the same via the NVIDIA app. The following options will be available under “Driver Settings” for supported apps:
- DLSS Override for Frame Generation – Enables Multi Frame Generation for GeForce RTX 50 Series users when Frame Generation is ON in-game.
- DLSS Override for Model Presets – Enables the latest Frame Generation model for GeForce RTX 50 Series and GeForce RTX 40 Series users, and the transformer model for Super Resolution and Ray Reconstruction for all GeForce RTX users, when DLSS is ON in-game.
- DLSS Override for Super Resolution – Sets the internal rendering resolution for DLSS Super Resolution, enabling DLAA or Ultra Performance mode when Super Resolution is ON in-game.
