It's been a long time since multi-GPU technologies like NVIDIA's "SLI" and AMD's "CrossFire" became things of the past. However, within the PC gaming community, a new technique for reviving old graphics cards in modern gaming environments is now drawing attention.
The system requirements of the latest AAA titles continue to balloon. With the spread of real-time ray tracing and high-resolution textures, the load placed on GPUs has become incomparable to that of previous generations. In response, companies have been rolling out AI-based frame generation technologies one after another. However, the benefits of this technology have often been limited to owners of the latest-generation graphics cards.
Reddit user quziwuzzi achieved smooth 144 FPS gameplay at 4K resolution (3840x2160) using an unusual combination of a high-end "GeForce RTX 3090" and an entry-level "GeForce RTX 3050". Behind this achievement lies not a reconstruction of multi-GPU architecture, but a complete division of labor between roles.
Dividing Rendering and Frame Generation
At the core of this technique is a utility software called "Lossless Scaling", sold on Steam for about $7. It originally started out as a simple upscaling tool that stretched old games—which could only run in windowed mode—to fullscreen without blurring the image. However, it has since evolved into a general-purpose tool for enhancing image quality and frame rates, equipped with its own machine-learning-based frame generation feature (LSFG) that doesn't depend directly on any specific game engine.
Lossless Scaling differs from native implementations like NVIDIA's DLSS or AMD's FSR, which integrate directly into a game engine's internal pipeline. Instead, it works by reading the completed 2D image rendered on screen, analyzing motion vectors within that image, and performing AI upscaling and frame interpolation as a post-processing step. Thanks to a recent update, Lossless Scaling now has the ability to completely offload the heavy computational burden of frame generation to a physically separate GPU from the main one.
In quziwuzzi's setup, the primary GPU—a Gainward RTX 3090—dedicates 100% of its resources solely to the game's 3D rendering. Meanwhile, the secondary GPU—an ASUS RTX 3050 (6GB model)—handles only the frame generation (interpolation) processing via Lossless Scaling.
Normally, even native frame generation consumes roughly 10% to 20% of GPU resources for the process itself. By offloading this to a secondary GPU, the primary GPU can operate without sacrificing any of its native rendering performance. Performance that was capped at around 71 FPS at 4K resolution with the RTX 3090 alone doubled to 144 FPS once assisted by frame generation on the RTX 3050. This successfully allowed quziwuzzi to fully unlock the refresh rate potential of their 144 Hz 42-inch OLED TV, the "LG OLED C4".
Why It Works Without SLI
Legacy multi-GPU technologies such as SLI and CrossFire relied on methods like "Alternate Frame Rendering (AFR)", where multiple GPUs took turns rendering frames. This required tight driver-level coordination and, in principle, demanded that users pair GPUs of identical models (or at least matching VRAM capacity and architecture). Furthermore, these technologies suffered from deep-rooted issues such as micro-stuttering caused by misaligned frame output timing, and were ultimately abandoned by both manufacturers and game developers as not worth the development cost.
However, this new technique requires neither a physical SLI bridge cable connecting the hardware nor two identical GPUs. From the OS's perspective, the two GPUs are simply recognized as independent display adapters, with the software asynchronously utilizing their respective computing resources.
That said, there is one clear hardware requirement: the motherboard must properly support PCIe lane bifurcation.
quziwuzzi used an ASUS "X870E ProArt Creator Wi-Fi" motherboard, configuring both graphics cards to operate in PCIe 4.0 x8 mode via BIOS settings. Transferring the completed frames rendered by the main GPU to the secondary GPU at high speed, then compositing intermediate frames there before outputting the final image, requires sufficient bandwidth and lane bifurcation on the motherboard's part. It's also worth noting that the display output cable itself must be connected to the secondary GPU (the RTX 3050, in this case), which generates and sends out the final displayed image.
Additionally, frame generation doesn't unconditionally improve performance. Even when rendering computation is offloaded to a separate GPU, the entire process of capturing, analyzing, generating interpolated frames, and outputting the image inherently introduces physical input lag. However, because the base frame rate was sufficiently high at 71 FPS, the perceived latency remained minimal even when using the 2X frame generation mode.
A New Use for Surplus Hardware
What these test results suggest is the potential for old, retired GPUs—or cheap, low-end GPUs—to gain new value as "co-processors".
Currently, the GPU market remains heavily affected by excess demand for data center chips driven by the generative AI boom, as well as memory shortages, keeping prices for the latest graphics cards elevated. AI frame generation via DLSS 3 is treated as a privilege exclusive to the RTX 40 series and later, making the financial barrier to upgrading extremely high for owners of older models. But with Lossless Scaling, this kind of manufacturer-imposed hardware restriction can be circumvented.
Simply pulling an old graphics card out of a drawer and plugging it into an open slot in your system allows it to function as a dedicated frame generation accelerator. As long as you have a motherboard that supports PCIe lane bifurcation, you can achieve smoothness approaching that of a high-end machine using a combination of existing hardware—without having to buy the latest GPU costing well over a thousand dollars.
On the other hand, this configuration also brings new challenges related to power supply capacity and heat dissipation. Since two GPUs run simultaneously, the overall system power consumption reliably increases compared to running a single GPU, and greater attention to case airflow becomes necessary.
User ingenuity has always evolved by threading through the gaps left in the roadmaps laid out by manufacturers. Will this new use for dual GPUs become the next trend among PC gamers hesitant to upgrade? Next, the community will likely test the limits of PCIe bandwidth and the practicality of latency when incorporating even older GPU architectures—such as the GTX 700 series or the 10 series.