A developer has shared a project called Neural Coprocessor on GitHub, as reported by Tom's Hardware. The project explores using two graphics cards in a novel way: one card runs a game normally, while the second card handles DLSS 5 neural rendering. DLSS (Deep Learning Super Sampling) is a technology from Nvidia that uses AI to enhance image quality or performance in games. In tests using a GeForce RTX 5060 Ti, the game ran at about 127 frames per second without neural rendering. When the same rendering was done on the same card, performance dropped to 59 frames per second. However, when the rendering was moved to a second card, performance improved to 106 frames per second.
The project uses a module called MGPU Bridge, which is an add-on for ReShade, a popular tool that lets gamers tweak graphics in real-time. The developer clarified that this is not related to SLI (Scalable Link Interface), an older technology that split the workload of rendering a single image across multiple cards. In this case, the neural rendering is a final step that takes a completed image and produces a new one, meaning it can be done on a separate card without interfering with the primary rendering process. This concept is reminiscent of Nvidia’s older PhysX cards, which were dedicated to specific tasks like physics calculations.
The tests were conducted using two RTX 5060 Ti cards with 16 GB of memory each. Neural rendering always operates at the display resolution, so performance gains are limited when using lower DLSS quality settings. On the card doing the rendering, the cost of processing ranged from 36% to 59% of the image processing workload, making upscaling less effective. When the rendering was offloaded to the second card, the cost dropped significantly to 0% in DLAA mode and 17% at worst. The primary card also ran 21 degrees Celsius cooler, as the workload was no longer concentrated on a single card’s cooling system.
While the results are promising, the project is still in early stages. The code is primarily research-based, tested on only two machines for up to twenty minutes at a time. It also requires two RTX 50-series cards, two monitors, and a DirectX 12-compatible game — a rare setup, especially since DLSS 5 is currently only supported in one game, NBA 2K27. However, the demonstration shows potential for improving performance and cooling, giving Nvidia a new angle to consider as it prepares to refine DLSS 5 ahead of its release on the GeForce RTX 40 series.
Neural Rendering Technique Demonstrates Performance Benefits Using Dual GPUs
AI-rewritten from original reportingHow it works
dlss5mgpu-bridgenvidiagaminggpu-renderingreshade



