← Back to Articles Explainer

DLSS vs FSR: How GPU Upscaling Technologies Work and Which to Use

DLSS and FSR serve the same user goal—higher frame rates at lower GPU cost—but the underlying methods differ substantially. DLSS uses a trained neural network running on dedicated tensor hardware. FSR uses spatial or temporal reconstruction that runs on any GPU. The tradeoffs in image quality, input latency, and hardware lock-in follow directly from that architectural difference.

Published June 2026 — VoltGround

Both DLSS and FSR reduce the number of pixels the GPU renders per frame and then reconstruct a higher-resolution output image. Running at 1080p internally to produce a 4K output means the GPU is shading approximately one-quarter of the pixels, which scales frame rates dramatically for GPU-bound scenarios. What differs between the two technologies is how the reconstruction step works and what hardware is required to do it.

DLSS: Tensor Core Neural Network Upscaling

DLSS (Deep Learning Super Sampling) is NVIDIA’s proprietary implementation. From DLSS 2 onward, the core of the technology is a convolutional neural network trained offline on high-resolution image pairs. At runtime, DLSS takes the current low-resolution frame, the previous frame, and motion vectors from the game engine, then feeds these through the network to produce the upscaled output.

The network inference runs on Tensor Cores, which are dedicated matrix multiply accelerator units present on all RTX 20, 30, and 40 series GPUs. On cards without Tensor Cores—GTX 10 and GTX 16 series, all AMD GPUs—DLSS is unavailable regardless of driver or game support. This is a hard hardware lock, not a software limitation.

Because DLSS uses temporal information (the previous frame and motion vectors), it reconstructs fine detail and thin geometry better than purely spatial approaches. The network has been trained to handle subpixel features and high-frequency textures. In practice, DLSS Quality mode output at 4K is frequently indistinguishable from native 4K rendering in static screenshots and often sharper in motion due to temporal stabilization.

FSR: Spatial and Temporal Upscaling Without Hardware Lock-in

AMD’s FidelityFX Super Resolution (FSR) takes a different approach to hardware requirements: it runs entirely on standard shader compute hardware and works on any GPU that supports DirectX 12 or Vulkan. This includes NVIDIA GPUs, Intel Arc, and even integrated graphics.

FSR 1 used a purely spatial algorithm (Lanczos-based edge reconstruction) without any temporal component. It was fast to compute but produced noticeably lower quality on thin geometry and text compared to DLSS 2 at equivalent quality modes.

FSR 2 introduced a temporal upscaler with motion vector support, substantially closing the quality gap with DLSS 2. FSR 2 and later versions reconstruct detail across frames similarly to DLSS, though without the trained neural network component. In most titles, FSR 2 Quality mode produces results that are competitive with DLSS Quality mode and clearly better than FSR 1 at the same resolution scale.

Hardware coverage: FSR runs on every modern GPU regardless of vendor. If a game supports both DLSS and FSR, NVIDIA RTX users should prefer DLSS for slightly better quality. Non-RTX and AMD GPU users have FSR as their primary temporal upscaling option.

Quality Modes and Internal Resolution

Both technologies offer multiple quality presets that set the internal rendering resolution as a fraction of the target output resolution. Using 4K output as the reference:

Ultra Quality / Quality: Renders at approximately 67% of output resolution (about 2560×1440 internal for 4K output)
Balanced: Renders at approximately 58% (about 2214×1244 internal)
Performance: Renders at 50% (1920×1080 internal for 4K output)
Ultra Performance: Renders at approximately 33% (1280×720 internal for 4K output)

Ultra Performance mode produces the highest frame rate gains but at the cost of significant reconstruction artifacts, particularly on thin objects, hair, and fast-moving detail. For competitive gaming where frame rate matters more than image quality, Performance or Ultra Performance is usable. For single-player titles where image fidelity is prioritized, Quality or Ultra Quality mode produces the best result.

Frame Generation: DLSS 3 and FSR 3

DLSS 3 (available on RTX 40 series only) introduced Frame Generation, which inserts a fully AI-synthesized frame between each rendered frame. This is distinct from upscaling: Frame Generation does not reduce rendering cost per frame but instead doubles the visible frame rate by generating intermediate frames using optical flow analysis on the Optical Flow Accelerator hardware present only in Ada Lovelace (RTX 40) GPUs.

Frame Generation increases the displayed frame rate counter significantly but does not reduce input latency in proportion—because the generated frames are synthesized, not rendered with new game state input. NVIDIA’s Reflex technology is recommended alongside Frame Generation to compensate for the latency offset. For single-player and racing titles, the visual smoothness improvement is real. For fast-twitch competitive play, native frame rate with Reflex enabled typically feels more responsive than Frame Generation at the same displayed number.

FSR 3 Frame Generation works differently: it does not require dedicated optical flow hardware and runs on standard shader compute. It is therefore available on any GPU that supports FSR 3 in a given game, including NVIDIA RTX 30 series cards that cannot use DLSS 3 Frame Generation. The quality of synthesized frames from FSR 3 is generally considered slightly below DLSS 3 Frame Generation in artifact-prone scenarios, but it is functional and widely available.

Choosing Between DLSS and FSR

For RTX 20 and 30 series users, DLSS 2 Quality mode outperforms FSR 2 Quality mode in image quality benchmarks for most titles, but the difference is small enough that either is acceptable. Use DLSS when the game supports it and your GPU is RTX; use FSR on AMD or older NVIDIA hardware.

For RTX 40 series users, DLSS 3 with Frame Generation and Reflex delivers the highest effective frame rate for single-player titles at minimum input latency cost. It is one of the more compelling hardware-exclusive features in recent GPU generations.

For AMD RX 7000 and RX 9000 series users, FSR 3 with Frame Generation is the equivalent path. The hardware support is broader but the quality ceiling of the frame generation synthesizer is slightly lower in demanding scenarios.