Why upscaling and frame generation do not work on all video cards

Upscaling and frame generation technologies have become an integral part of gaming graphics rendering over the past five years. They allow for high image quality at a lower performance cost, and improve the smoothness of gameplay even at ultra settings. However, most of these solutions — such as NVIDIA's DLSS, AMD's FSR, and Intel's XeSS — are limited in compatibility: some only work on certain generations GPU, others are exclusive to one architecture. Why is this? In this article, we will take a detailed look at the origins, development, and hardware limitations of the most popular upscalers to understand why they are not supported on all video cards.

How Upscaling Came to Be: Using DLSS and DLSS Frame Generation as an Example

Modern upscaling is the result of a long evolution of rendering technologies, starting with simple filtering and anti-aliasing and leading to the use of neural networks and machine learning. The turning point came in 2018, when NVIDIA introduced the first version of DLSS (Deep Learning Super Sampling). Although the technology was initially planned as an intelligent anti-aliasing algorithm, its creators, led by Jensen Huang, quickly realized that the potential of DLSS goes far beyond simply combating “jaggies” at the edges of objects. The bet was made on increasing performance by upscaling an image from a lower resolution to a higher one while maintaining visual quality comparable to native rendering.

DLSS 1.0 used game-trained neural network models, which required significant effort from developers: NVIDIA had to analyze each game in advance and train the model based on the unique characteristics of the scene and camera behavior. This limited flexibility and led to quality complaints: some titles showed excessive blurriness or artifacts.

DLSS 2.0 completely changed the approach: a universal architecture was created that uses temporary buffers, motion data, depth and color maps, as well as an improved algorithm for detail approximation. This allowed for significant quality improvements and simplified integration into games. DLSS 2.0 worked stably and efficiently on many projects without additional training.

With DLSS 3, NVIDIA went further: frame generation was introduced (Frame Generation), in which the system not only improved the resolution, but also created new frames based on the analysis of the motion between the previous two. This was made possible by the Optical Flow Accelerator, a block capable of interpreting motion vectors and scene features at the level of pixel streams. However, full-fledged frame generation requires not only a good data flow, but also fast decision-making in real time. That's why DLSS 3 became an RTX 40 exclusive: the Ada Lovelace architecture made it possible to achieve the required processing speed thanks to the third generation of Tensor Core and the Reflex system.

The next step, DLSS 4, arrived in 2025 and was exclusive to the RTX 50. In this version, NVIDIA introduced Multi Frame Generation — a method in which several “predicted” frames are inserted between each real frame. This became possible thanks to the use of transformer neural network models — the same ones used in modern NLP and generative AI. Such models require enormous computing power and bandwidth, as well as specialized hardware units for training and inference. RTX 50 was the first to offer such components: fifth-generation Tensor Core, accelerated RT units, and expanded caches. Thus, DLSS evolved from anti-aliasing to a full-fledged video generator, but each step forward was accompanied by increased hardware requirements.

How FSR and FSR frame generation came about

AMD took a completely different path. In 2021, the company introduced the first version of FidelityFX Super Resolution (FSR), starting with the most universal solution: FSR 1.0 was a fully spatial upscaling algorithm. It did not rely on temporal information, did not require machine learning, and ran on any modern GPU, including NVIDIA cards and even integrated Intel graphics. This approach ensured the technology's immediate adoption, but the quality was lower than that of DLSS.

With FSR 2.0 in 2022, AMD added temporary buffers and motion vectors to the algorithm. This gave a quantum leap in upscaling, making the image less noisy and increasing stability in dynamics. However, FSR still did not use neural networks - its main goal was accessibility.

This changed with the release of FSR 3 in 2023: AMD introduced frame generation, similar to DLSS 3. And although the technology remained “open”, FSR 3 required higher performance to work correctly, including efficient implementation of optical flow and work with temporary buffers. For this reason, frame generation in FSR 3 did not work at all. GPU.

FSR 4, released in 2025, became even more demanding. In this version, AMD added machine learning elements for the first time: local neural network filters, temporal motion models, and camera behavior prediction are used. This required new blocks within the RDNA 4 architecture, unofficially called AI accelerators. These components are only present in the Radeon RX 9000 and newer. So, while the basis of FSR 4 remains open, its advanced FG (Frame Generation) work exclusively on the latest AMD video cards.

How XeSS came to be

XeSS (Xe Super Sampling) was introduced by Intel in 2022 as a response to existing solutions from AMD and NVIDIA. Unlike its competitors, Intel initially set a goal to make its technology universal and cross-platform. XeSS uses machine learning, but implements it taking into account differences in hardware. Intel Arc graphics cards use hardware acceleration through XMX blocks (Xe Matrix Extensions) - specialized modules similar to NVIDIA's Tensor Core. These blocks are designed to quickly perform matrix operations required for neural network inference.

For other video cards, XeSS supports a fallback mode based on DP4a instructions, which is a feature of accelerated calculations via SIMD, present in GPU starting with NVIDIA GTX 10-series and AMD RDNA 1. However, image quality and performance are inferior to XMX mode. Such flexible architecture allowed Intel to provide minimal compatibility with a wide range of hardware.

Intel promotes XeSS as an open-spec technology. However, the latest updates — versions 1.3 and higher — have added support for experimental frame generation functions. So far, they are only available on Intel Battlemage graphics, which implement the second version of the XMX cores. These cores have improved throughput, support for variable precision (FP8/INT4), and optimization for real-time tasks.

Thus, although XeSS started as a maximally open and compatible project, its evolution shows that advanced features are also starting to require hardware support. And while basic upscaling is available on almost all GPU, then features similar to DLSS 3 and 4 require the latest Intel GPUs.

Why DLSS 1 and 2 weren't supported on non-RTX graphics cards

The limitation of DLSS 1 and 2 only on RTX graphics cards is explained by the technical architecture of these solutions. Both versions were initially developed based on Tensor Core — specialized modules for neural network operations, first introduced in the Turing architecture. These cores provide high speed of matrix multiplication, which is necessary for performing convolutional operations in deep neural networks.

GTX series (e.g. GTX 1080 Ti) and AMD graphics cards do not have such units. Even if you try to emulate DLSS using regular CUDA cores or shader units, the performance drops by tens of times, and the result becomes useless. In addition, DLSS uses NGX SDK (Neural Graphics Acceleration), which checks for Tensor Core and RTX architecture at the driver level.

Thus, even with a powerful graphics chip (for example, TITAN V), it is impossible to run DLSS: the lack of NGX integration and Tensor Core blocks access to the API. Third-party attempts to modify the SDK or run DLSS through external hacking methods do not produce results, since the logic of DLSS itself is based on a tight connection between hardware and software code.

Why DLSS 3 frame generation is supported on video cards not lower than RTX 40

Frame generation in DLSS 3 relies on Optical Multi Frame Flow technology, which is implemented in the Ada Lovelace architecture. To analyze the movement between frames, the third-generation Optical Flow Accelerator is used - a specialized hardware module that is missing from the RTX 30 and younger series.

This accelerator calculates motion vectors between pixels in two frames using depth maps, velocity buffers, and shadow masks. Without this module, it is impossible to accurately generate an intermediate frame - the result will be visual artifacts and "ghosting" of the image. This is why the RTX 30, despite the presence of Tensor Core, cannot support DLSS 3 Frame Generation.

Additionally, DLSS 3 requires interaction with the NVIDIA Reflex system, which synchronizes frames between the CPU and GPU, reducing input lag. Without Reflex, introducing additional frames would greatly increase control latency. Reflex support in DLSS 3 is strictly tied to the Ada Lovelace architecture, making it impossible on RTX 30.

Why DLSS 4 frame generation is supported on video cards not lower than RTX 50

DLSS 4 requires Blackwell architecture and the new fifth-generation Tensor Core. This version uses transformer models — complex neural networks that can predict scene dynamics 3–4 frames ahead. This requires buffering multiple temporal layers: optical flow, depth maps, motion masks, particle states, and camera behavior.

This multi-level processing requires high-speed access to video memory, an additional cache line, and a redesigned ALU hierarchy. All of these components are first implemented in RTX 50. Even RTX 40, with its powerful architecture, cannot handle 4-6 temporary buffers simultaneously in real time.

Additionally, DLSS 4 requires double the bandwidth for mixed-mode operations (FP16/INT8) and the presence of transform units for adaptive scalability - these features are built into Tensor Core 5.0. Thus, the shift towards Multi Frame Generation required a radical architecture update, which makes DLSS 4 impossible on older cards.

Why FSR frame generation is supported on video cards not lower than RX 9000

AMD's FSR 4 is the first to implement ML-based adaptive frame generation. Unlike FSR 3, which generated frames using a hard-coded algorithm based on optical flow, FSR 4 uses a trainable prediction system based on temporal patterns and analysis of previous scene states. These tasks require hardware AI accelerators — AI Compute Units — which were introduced in the RDNA 4 architecture.

The RX 6000 and 7000 series (RDNA 2 and 3) graphics cards do not contain these units. In addition, they lack the memory bus width required for parallel processing of motion buffers and predictive models. FSR 4 also uses an updated FidelityFX SDK version 5.0, which is incompatible with the outdated GCN and RDNA 2 driver microcode base.

FSR 4 requires at least 64 AI cores, BFLOAT16 support, variable instruction length, and INT4 processing - all of which were introduced in the RX 9000. So, despite the formal openness of FSR, the new frame generation works exclusively on the latest AMD graphics cards.

Conclusion

Modern upscalers are no longer simple image stretching algorithms. These are complex systems that include elements of computer vision, motion analysis, work with temporary buffers, and even transformer neural networks. Therefore, it is not surprising that specialized hardware is required for their full operation: tensor cores, optical flow accelerators, AI blocks, and advanced caches. Each new iteration of DLSS, FSR, or XeSS raises the bar, but also more tightly ties the technologies to specific generations of video cards. Thus, the lack of support on old GPU is not due to the greed of manufacturers, but to objective technical limits. Progress requires new solutions - and new chips capable of handling the increased complexity of graphics in real time.