Deep Learning Super Sampling (DLSS) is NVIDIA’s technique for rendering a game at a lower internal resolution and using a neural network to reconstruct a sharp, high-resolution image - trading a small amount of AI inference for a large gain in frame rate. First shipped with the RTX 20-series GPUs in 2018, it matured dramatically with DLSS 2.0, announced on NVIDIA’s developer blog in March 2020.
According to that official post, the network is “trained on tens of thousands of high-resolution, beautiful images, rendered offline in a supercomputer at very low frame rates and 64 samples per pixel.” The trained model then runs in real time on the GPU’s Tensor Cores - “up to 110 teraflops of dedicated AI horsepower” on Turing - alongside the game itself. DLSS 2.0 added “new temporal feedback techniques for sharper image details and improved stability from frame to frame,” letting it reconstruct “up to 4X super resolution (i.e. 1080p to 4K)” while the game renders “one quarter to one half of the pixels.”
DLSS is a landmark because it put deep learning directly into the real-time graphics pipeline of mainstream games, where milliseconds matter. It made high-resolution, high-frame-rate play feasible on more modest hardware and kicked off an industry-wide wave of AI upscaling, with competing approaches following from other GPU makers and platforms.