Imagine being handed someone else's construction blueprint and asked to pour an entirely new concrete wall—no formwork, no rebar, only the final shape. That is the software version of what a project circulated on September 20, 2026 does: rewriting NVIDIA's DLSS 5 neural rendering from scratch and getting it running on an Intel integrated GPU.
Not a Wrapper, a Full Re-Pour
The project, authored by Uzbekunknown under the name dlss-nr-on-intel, targets the Arc 140V integrated GPU inside an Intel Core Ultra 7 256V (Lunar Lake, Xe2 architecture, eight Xe cores, up to 64 TOPS INT8). The system is Arch Linux with Mesa drivers.
The pipeline has three key design choices:
1. The network itself. The project does not wrap or bridge NVIDIA's dynamic libraries. It fully rebuilds the 71-block U-Net used by DLSS 5 neural rendering—which is why it needs roughly 2.3 GB of device buffers. The network actually runs; nothing is forwarded.
2. Precision. NVIDIA stores weights in FP8, which Xe2 does not support, so weights are converted to FP16 while accumulation stays in FP32 for numerical stability.
3. The interface. Inference runs via Vulkan's VK_KHR_cooperative_matrix extension on the XMX matrix engines—no CUDA, no NVIDIA NGX. It hooks vkQueuePresentKHR, so in theory it works with any Vulkan-presented application.
Timing context: DLSS 5 neural rendering launched on September 4, 2026. In the following weeks, the community ported it to older RTX cards, then AMD's RX 9000 series, and now Intel integrated graphics—the technical barrier keeps dropping.
Running vs. Usable: 412 Milliseconds Apart
Performance data leaves no room for illusion:
- 640x360 (about one-eleventh the area of 1080p): ~10.5 FPS in Tekken 7
- 720p: 3-5 FPS, with device buffers consuming ~2.3 GB of shared system memory
- 1080p: 1.5-2 FPS; a single neural rendering inference takes ~412 ms, capping the theoretical frame rate at ~2.4 FPS even if game rendering were instantaneous
- When a Windows-compatible build will appear
- Actual frame rates on desktop-class Intel discrete GPUs
- Baseline results on AMD RDNA3 and newer architectures
- The provenance boundaries and licensing status of the weights
- Whether third parties can independently reproduce the same numbers with AI-written porting code—and whether they stumble into the same nonexistent "driver defect"
This is a runnable proof of concept, not a usable solution. The author notes that desktop-class Intel discrete GPUs (like the A770 or B580) would deliver significantly better frame rates thanks to higher memory bandwidth.
Image Quality: The Model's Taste Exposed
Validation used three fighting games: Dead or Alive 5 Last Round, Tekken 7, and Mortal Kombat 1. Since the DLSS 5 NR model was trained on photorealistic imagery, detail enhancement and noise suppression worked well on Tekken 7 and Mortal Kombat 1.
On the more anime-styled Dead or Alive 5 Last Round, the model over-enhanced textures and lighting, giving characters' faces an unintended "aged" look—worse than native rendering overall. A model trained on realistic images actively "corrects in the wrong direction" on stylized content: it pushes hard-lined anime faces toward photorealism rather than simply blurring or sharpening. This is a problem that must be solved before such technology enters stylized games, animation, or fan-content pipelines.
The Code Was Mostly Written by Two Models
The most counterintuitive part: the author says the code was written primarily by Anthropic's Claude and OpenAI's GPT-6 Astra. He supplied the hardware platform, original binaries, and high-level technical direction; modeling, adaptation, and debugging were executed autonomously by AI agents.
The project also preserves a rarely disclosed record: while analyzing driver-layer behavior, the AI generated a nonexistent hypothesis of a "driver-level compatibility defect." This fabricated issue was treated as a real obstacle and influenced decision-making across three key development stages. The author kept this detour in the project documentation as a record of cognitive bias.
The striking part is how long the false hypothesis survived—it wasn't dismissed as casual hallucination but entered engineering decisions across three phases. Breaking that chain requires either someone re-verifying assumptions mid-course, or a "falsify before proceeding" gate in the workflow. This is the most transferable lesson of the event.
Where It Sits Among Post-Processing Pipelines
This must be distinguished from another route: AMD's ECCV demonstration of single-step latent diffusion for indirect illumination, where the GPU computes direct light and a diffusion model fills in diffuse indirect lighting. That approach is pipeline-embedded—its input is the engine's 3D data.
DLSS 5 neural rendering, described as 3D-Guided Neural Rendering, performs deep scene reconstruction in post-processing on finished frames; its inputs are rendered frames and motion vectors. The former needs engine cooperation; the latter only needs the picture—which is why this port could bypass the engine entirely by hooking the presentation layer.
A side effect: users must supply their own DLSS files to extract weights. Important caveat: this is a community project, not endorsed or supported by NVIDIA, and its terms-of-use and licensing status are undisclosed.
What Remains to Be Verified
A ten-year-old Tekken 7 running at 10.5 FPS at one-eleventh resolution is a useless number on its own. What it proves is something else: a neural network locked inside a specific vendor's execution library can be re-poured from scratch by outsiders and run on entirely different matrix units. The next question: when porting someone else's model to your own hardware costs one person plus two models, does the hardware vendor's moat still live in the silicon—or has it moved somewhere else?
---
References
1. VideoCardz, *Modder gets DLSS 5 Neural Rendering running on Intel Arc 140V integrated graphics*, 2026-09-20. https://videocardz.com/newz/modder-gets-dlss-5-neural-rendering-running-on-intel-arc-140v-integrated-graphics 2. Wccftech, *Intel Arc 140V Runs DLSS 5 Neural Rendering Without CUDA, But Frame Rates Collapse To Just 3-5 FPS*, 2026-09-20. https://wccftech.com/intel-arc-140v-runs-dlss-5-neural-rendering-without-cuda 3. Mydrivers / Sina Tech, *DLSS 5 ported to Intel Arc integrated graphics, written mostly by AI: Arc 140V hits 10.5 FPS*, 2026-09-20. https://k.sina.com.cn/article_5952915720_162d2490806704va9k.html 4. PC Central, *Intel Arc 140V Runs DLSS 5 Neural Rendering in Linux Proof-of-Concept*, 2026-09-20. https://pccentral.net/intel-arc-140v-runs-dlss-5-neural-rendering-linux-proof-of-concept 5. GameGPU, *DLSS 5 works on Intel Arc Xe2 at 720p with performance up to 3-5 FPS*, 2026-09-20. http://en.gamegpu.com/news/zhelezo/dlss-5-zarabotal-na-intel-arc-xe2-v-720p-s-proizvoditelnostyu-do-3-5-fps