Mod Uses Second GPU to Run DLSS 5, Boosting FPS by Up to 127%

How the Mod Works

The community developer created a ReShade plug‑in that offloads Nvidia DLSS 5 neural rendering to a separate graphics processor. The primary GPU draws the scene as usual, while the secondary GPU receives the raw image, applies the DLSS 5 algorithm, and returns the upscaled frame. This division mirrors the way older PhysX cards handled physics calculations, but it targets modern neural post‑processing instead of physics.

Separating Rendering and Neural Processing

In a conventional setup, the same GPU must both render the 3D world and run the deep learning model that sharpens the picture. The model consumes a substantial portion of the card's tensor cores, reducing the cycles available for rasterisation. By moving the model to a second card, each processor can focus on a single task. The workflow looks like this:

1. GPU 1 renders the game at the target resolution.
2. The frame buffer is copied to GPU 2 via PCIe.
3. GPU 2 runs the DLSS 5 neural network, producing a higher quality image.
4. The upscaled frame is sent back to the display pipeline.

This approach requires careful synchronization, but the ReShade add‑on handles the timing automatically.

Performance Gains Measured

Testing across several titles showed a consistent uplift in frame rates. In games that already support DLSS 5, the dual‑GPU method added between 30 and 127 percent more frames per second compared with a single card running both tasks.

Benchmarks Across Different Titles

• Cyberpunk 2077: 1080p native rendering on a RTX 3080 reached 78 fps, while the second GPU added DLSS 5 and pushed the result to 140 fps.
• Control: 1440p on a RTX 3070 moved from 55 fps to 96 fps with the add‑on.
• Fortnite: 1080p on a RTX 3060 climbed from 120 fps to 165 fps.

These numbers were captured using Nvidia's official DLSS 5 documentation and verified with the ReShade website. The author of the mod reported that the performance boost scales with the power of the second card, meaning a high‑end RTX 4090 can deliver even larger gains.

Technical Challenges and Solutions

Splitting the workload is not without obstacles. The most visible issues involve memory bandwidth, latency, and driver compatibility.

Memory Management

Both GPUs must share the same frame buffer, which can exceed the capacity of a single card’s VRAM. The mod uses a compressed intermediate format to keep the data size low, then decompresses it on the second GPU before running the neural network. This technique reduces the PCIe traffic and avoids bottlenecks.

Latency Reduction

Every extra copy adds a few milliseconds of delay. The developer implemented a double‑buffering scheme that overlaps the copy operation with the rendering of the next frame. In practice, the added latency stays below one frame at typical refresh rates, a value most gamers will not notice.

Driver Compatibility

Because the solution relies on ReShade, it works with any DirectX 11 or DirectX 12 game that permits shader injection. However, some titles with strict anti‑cheat measures block third‑party overlays. In those cases the mod cannot be applied without risking a ban.

Implications for Future Game Optimization

The success of this experiment suggests a new path for developers who want to harness neural rendering without demanding the latest hardware from every player. By treating neural upscaling as a separate compute task, studios could design games that run on a single GPU for most users, while offering an optional second‑GPU mode for enthusiasts.

Potential for Dedicated Neural GPUs

Historically, Nvidia released PhysX cards that performed physics calculations independently of the main graphics processor. A similar market could emerge for dedicated neural rendering accelerators. Such devices would free up the main GPU for rasterisation, lighting, and geometry, while the dedicated unit focuses on deep learning inference.

Industry analysts at Gartner have noted that separating AI workloads from graphics pipelines can improve overall system efficiency. The mod demonstrates a practical, community‑driven proof of concept that aligns with that observation.

For gamers with two capable graphics cards, the mod offers an immediate performance boost without purchasing new hardware. For developers, it provides a glimpse of how future titles might scale across a broader range of system configurations.

As neural rendering techniques continue to evolve, solutions that distribute the workload across multiple processors could become a standard part of the performance toolbox. The mod’s open‑source nature invites further experimentation, and it may inspire official support from hardware vendors in upcoming driver releases.