Overview
- Nvidia’s DLSS upscaling shifts from CNNs to a transformer-based model that assesses each pixel for more precise image reconstruction.
- The transformer model became widely available to developers with the release of DLSS SDK 310.3.0, marking the end of its beta phase.
- Doubling the model’s parameters enables a 30–50 percent increase in ray tracing throughput and cuts processing latency to around 1 millisecond on an RTX 5090.
- All RTX GPUs from the Turing generation through Blackwell will support the new model, with RTX 50-series cards gaining exclusive multi-frame generation of up to three synthetic frames per real frame.
- Gamers can expect noticeably sharper visuals, reduced ghosting, and fewer artifacts when DLSS4 officially launches later this year.