Trending: Silksong patch 4Search
iHeartGeek
iGaming

DLSS 5 runs on Intel graphics thanks to AI agents

A research port runs Nvidia's neural rendering pass on an Intel Arc 140V integrated GPU, with the code written almost entirely by AI agents.

A graphics card cooler photographed in close-up under cyan and magenta light, its fan and fins filling the centre of the frame

A developer publishing on GitHub as Uzbekunknown has got Nvidia's DLSS 5 neural rendering running on an Intel Arc 140V, the integrated GPU inside a Lunar Lake laptop, under Linux. The project, dlss-nr-on-intel, is an independent reimplementation of the pass rather than a wrapper around Nvidia's own code, and it needs no Nvidia hardware, no NGX and no CUDA. The reimplemented network runs on Intel's XMX matrix units through the VK_KHR_cooperative_matrix extension, and the pass is injected at vkQueuePresentKHR, so it attaches to anything that presents a frame with Vulkan, including a Windows game running under Proton. The author is careful to call it a research port rather than a product.

What it actually does

DLSS 5's neural rendering pass is a one-step diffusion model that re-renders a frame's detail, and the port reimplements its 71-block network from scratch. It has been demonstrated in Tekken 7, Dead or Alive 5 Last Round and Mortal Kombat 1, across Unreal Engine 4, an older Direct3D 9 title and a VKD3D-Proton translation layer, with side-by-side stills and measured texture gains over the game's own output. Users have to supply their own copy of Nvidia's DLL; none of Nvidia's binaries or trained weights ship with the code. Performance is the obvious cost. The author reports around 10 frames per second at 640x360 in Tekken 7, with frametimes described as abysmal, and suggests photo mode is the realistic use case for now rather than anything you would play.

The interesting part is who wrote it

The port was written almost entirely by AI. The repository states plainly that the breakdowns and analysis are written by AI, and the developer credits Claude Opus 5 and OpenAI's Astra, saying that they supplied the hardware, the binary and the direction and that they cannot defend the code line by line. One detail stands out: a driver bug hallucinated by one of the agents shaped three phases of the development work before anyone noticed. This is the latest in a run of unofficial DLSS 5 ports that have appeared within weeks of each other, taking the technology to RTX 30-series cards, to AMD's RDNA 4 graphics cards and even into a web browser.

Our opinion

The hallucinated driver bug is the most instructive line in the repository, and it is worth more than the 10 frames per second it sits next to. A model that invents a plausible-sounding API and then builds three phases of engineering on top of it is exactly the failure mode that makes people nervous about AI-written code, and the honest response is neither outrage nor awe, because the author says so out loud instead of hiding it. What you are looking at is a demonstration of leverage: one person, a laptop with integrated graphics and a pair of coding agents produced a working reimplementation of a proprietary rendering pass in days. What you are not looking at is a product, and the port's own author is the first to say so. The genuinely interesting question is what happens when the same tools are pointed at something with a budget and a shipping date, because the modding scene has already answered the smaller one.