There are plenty of reasons to consider an Nvidia graphics card over an AMD one, but their superior ray tracing performance and support for DLSS upscaling technology are the two compelling reasons that immediately stand out, making them look better. These two often hold hands to deliver a good experience overall, as DLSS helps you maintain high frame rates in games without putting too much pressure on the GPU, so it can do the heavy lifting to render ray-traced lighting and shadows. It is, however, crucial to understand what DLSS actually is and how it works to decide whether it's even worth enabling.

What does DLSS mean, and what does it do?

A deep dive into Nvidia's upscaling tech

An image showing a side-by-side comparison showing Cyberpunk 2077 with and without DLSS enabled.

DLSS, short for Deep Learning Super Sampling, is an AI-powered image-enhancing technology that's exclusive to Nvidia's RTX graphics cards. It's actually one of the earliest examples of AI technology that has proven itself to be both useful and decently widespread. The idea behind it is pretty simple: use the AI hardware inside Nvidia GPUs to make games look better and play at a higher framerate.

DLSS 1/2 uses AI to increase resolution, and DLSS 3 uses AI to increase resolution and to make new frames.

There are three versions of DLSS, and this is where things can get confusing. The first iteration of DLSS was introduced in 2019 (in a Battlefield V update) but has largely been replaced by DLSS 2, which came out in 2020 and introduced much better visual quality, turning DLSS from a kind of pointless feature to something you'd actually want to enable. DLSS 3 came out in 2022, adding AI-made frames (or frame generation) into the mix. Basically, DLSS 1/2 uses AI to increase resolution, and DLSS 3 uses AI to increase resolution and to make new frames.

An image showing the Zotac Gaming GeForce RTX 4070 Super GPU kept on a table.

All Nvidia cards branded with RTX support DLSS, but to varying degrees. At the time of writing, only RTX 40 GPUs on the Ada Lovelace architecture, like the RTX 4090, support DLSS 3's frame generation technology, even though all RTX GPUs have AI hardware. Additionally, DLSS is only available in select games, such as Cyberpunk 2077 and Hitman World of Assassination. Today, just over 300 games support at least one version of DLSS, and 36 of those games include support for both DLSS 1/2 and 3.

A combination of rasterization and Tensor cores

DLSS is an incredibly complicated and cutting-edge technology, so here's the short version of how it works. All RTX GPUs have traditional rasterization cores that render the game, but also Tensor cores that enable AI acceleration. The idea is that those Tensor cores can take the frames that the rasterization cores create and improve the image quality or even create brand-new frames. In order to get the best image quality, though, game-specific AI training is necessary. That's because games vary widely in art direction and graphics, and an AI trained in Minecraft wouldn't be great to use in, say, The Witcher 3, for example.

DLSS 1/2 (which only uses resolution upscaling) is a performance-boosting setting. For example, if you set your resolution to 1080p and enable DLSS, the GPU isn't rendering the game at 1080p and using the Tensor cores to make that 1080p look like 1440p. Instead, it's rendering the game at 720p (or another similarly low resolution) and using DLSS to increase the resolution to look like 1080p. The ideal result is that the game looks the same but with a much higher framerate.

Nvidia DLSS 3 frame generation in action.
Source: Nvidia

DLSS 3 is basically DLSS 2 but adds another step for frame generation. After rendering and upscaling two frames, the Tensor cores will then observe the difference between those two frames and guess what would have happened in between, as illustrated by the above image. Compared to DLSS 1/2, DLSS 3 can boost the framerate by about 50%.

DLSS 3.5 with Ray Reconstruction

DLSS 3.5 is essentially an extension of DLSS 3 with the inclusion of a new "Ray Reconstruction" feature. This new feature promises to improve the visual quality of ray-traced games by replacing the previously hand-optimized denoiser with an AI network trained by supercomputers. With the help of Ray Reconstruction, higher-quality pixels are automatically created and inserted between sampled rays.

A screenshot showing the Cyberpunk 2077 optimized settings benchmark.
Cyberpunk 2077 performance benchmark with GeForce Experience "Optimized" settings.

The in-game framerate remains largely unchanged due to this, but there's a significant improvement in the image quality overall. We've already seen how impressive this can look in games like Cyberpunk 2077 and Alan Wake 2.

DLSS 3.5 and Ray Reconstruction are available to all graphics cards of the GeForce RTX GPU, starting from the Nvidia GeForce RTX 20-series. It is, however, recommended that you use RTX 40-series GPUs for the best results, as a combination of the following features works together to deliver a good experience when DLS 3.5 is enabled:

  • Super Resolution
  • Frame Generation
  • Ray Reconstruction
  • Nvidia Reflex
A render showing a list of various DLSS versions and supported GPUs.

The drawbacks of DLSS and why it's not a silver bullet

If this all sounds too good to be true, you'd be right. DLSS isn't perfect, and there are many inherent drawbacks to the technology. The most obvious of these is that DLSS is limited to just a few hundred games, most of which came out after 2018. There are very few titles before that year that have DLSS, so it's a feature that's largely limited to the latest AAA games.

Another problem is that it can easily run into CPU bottlenecks. Depending on the CPU and game, lowering the resolution (or any graphically intensive setting) may not boost the framerate as expected, either because the CPU is overwhelmed or the game can't efficiently harness the CPU's power. If you're CPU bottlenecked, DLSS won't increase your framerate much, if at all, because it achieves that increased framerate by lowering the real resolution. You'll still see an upscaled image but without the extra frames.

If you're CPU bottlenecked, DLSS won't increase your framerate much, if at all.

The frame generation part of DLSS 3 isn't impacted by CPU bottlenecking but has two major problems of its own. AI isn't that great at duplicating UI elements like text and minimaps, and DLSS 1/2 gets around this by only letting the AI upscale the 3D elements in the game and applying the UI afterward. However, DLSS 3 with frame generation is forced to use a completely rendered frame, including the UI, and this causes the UI to flicker and occasionally be garbled or even unreadable. Things seem to have improved with DLSS 3.5, but you'll likely notice it occasionally.

There's an even bigger problem with frame generation, though. To create an AI-made frame, two rendered frames are required, one of which needs to come after the AI-made frame, otherwise you'd be seeing frames out of order. This creates a ton of additional latency because the GPU makes you wait longer to get the latest frame. The end result is that the framerate is much higher, but the latency stays the same, even though increasing the framerate normally lowers the latency. This means the game looks smooth but doesn't respond to your button presses as fast as you would expect.

Despite its drawbacks, DLSS is still a leader

The rendering pipeline for FSR 3.
Source: AMD

Although DLSS has problems (particularly DLSS 3), it's still the best image enhancement and performance-boosting technology for games and has been ever since it debuted in 2019. That's not for a lack of competitors, either. AMD launched FidelityFX Super Resolution (or FSR) in 2021 and Intel, alongside its Arc Alchemist GPUs, launched Xe Super Sampling (or XeSS) in 2022. DLSS arguably provides better image quality than FSR and XeSS, is present in more games (about 250 for FSR and 50 for XeSS), and offers a unique frame generation option.

However, FSR and XeSS have been catching up pretty quickly. FSR has only been out for a few years at the time of writing and is close to matching DLSS's level of supported titles. XeSS is hardly more than a year old, and it's in at least 50 games. Additionally, FSR is supported on GPUs dating back to 2016 and works on AMD, Intel, and even Nvidia branded cards. Frame generation isn't exclusive to just Nvidia GPUs either. AMD's Fluid Motion Frames technology can produce similar results, and it can work at a driver level.

None of the upscaling technology comes close to matching the general performance and fluidity of Nvidia's DLSS, though, and it remains to be seen if that changes in 2024 when AMD will likely released an updated version of FSR and its Fluid Motion Frames techonologies.