Monday, Nvidia announced the second iteration of its Deep Learning Super Sampling (DLSS) or DLSS 2.0, with Control and MechWarrior5: Mercenaries adopting the technology. Nvidia’s GeForce RTX 20-series is the current cutting edge for GPUs. Its raytracing capabilities is a key selling point and a lot of conversation has been generated on the technology, but DLSS is another apart of the RTX 20-series appeal, specifically the 2080, 2080 Super, and 2080ti.
DLSS is a key part of Nvidia’s Turing architecture as a part of a project to reach photo-realism with their GPUs. DLSS uses machine learning to tap into tensor cores to increase frame rates and increase the resolutions of the game being played. This is intended to compensate for the weaknesses of raytracing, as the lightning technology drops frame rates to render the dynamic system.
Tensor cores are a specialized type of core designed to process the mathematics needed to preform High-Performance Computing and deep learning. Tensor cores were first debuted with Nvidia’s Volta GPUs but Volta was never released for individual customers, being reserved for AI and animation companies. Nvidia’s RTX 2080s are the first commercial cards to have tensor cores, although they are a bit different from Volta’s cores and will not require FP16 precision to run.
Nvidia did have an earlier iteration of DLSS, now retroactively called DLSS 1.0, but that iteration was not very popular. DLSS often had mixed results, mostly with character hair would drop in resolution and getting a grainy look, and frame rates suffered. The largest barrier for DLSS 1.0 not only did it have to be implemented in each game, but DLSS 1.o had the be trained and fed data on each individual game for it to perform properly. Battlefield V and Metro: Exodus implemented DLSS 1.0 post-launch to mixed results.
Nvidia put DLSS 2.0 in beta almost immediately having launched the first iteration, with Control being the test benchmark. Nvidia’s launch of DLSS 2.0 fixes most of the issues with DLSS 1.0, with most important fix being that DLSS 2.0 being pre-trained off of data that can be used in most games and accumulates more data off of each frame to better predict the next frame. Although, DLSS 2.0 needs to be implemented on a per-game basis.
Nvidia claims that DLSS 2.0 can enhance the gaming experience at native resolution at half the pixels as DLSS 1.0 and the rest of the power would be dedicated to boosting the frame rates, or boost both the performance and resolution past the native experience. Since DLSS 2.0 uses the tensor cores more efficiently than 1.0, players do not have to suffer resolution limitations that DLSS 1.0 imposed.
DLSS 2.0 will offer three modes, 4X, 2X, and Balanced. 4X prioritizes performance and drastically boosts frames, 2X focuses on boosting visuals, and Balanced splits between frames and visuals. Developers have the option on which modes are integrated but 2X will be the standard minimum if DLSS 2.0 is added. DLSS 2.0 will need to be integrated into each game individually, but with the lack of training needed, the process is streamlined requiring the game to only have temporal anti-aliasing.
Control and MechWarrior5: Mercenaries are the current flagship for Nvidia to demonstrate DLSS 2.0, but it has already been quietly implemented in Wolfenstein: Youngblood and Deliver Us to the Moon. Control will officially receive a patch to change from DLSS 1.0 to 2.0 when Control’s new DLC The Foundation, launches.