DLSS 5 on Intel Arc: Neural Rendering works, but it's a slideshow
Following the emergence of independent implementations for AMD Radeon cards, the next milestone has arrived for Intel. A research project hosted on GitHub has successfully demonstrated Neural Rendering (the core technology behind DLSS 5) running on the integrated Arc 140V graphics. While technically significant, the current results offer a slideshow rather than a playable gaming experience.
DLSS 5 Without NVIDIA Components
The GitHub project, titled “dlss-nr-on-intel,” is a research port that implements Neural Rendering on the Xe2 architecture under Linux. Notably, this implementation does not rely on NVIDIA hardware, CUDA, or the NGX runtime. Instead, it utilizes the VK_KHR_cooperative_matrix Vulkan extension to access the XMX matrix units built into the Intel GPU.
However, the project is not entirely independent of NVIDIA's software; users must manually extract the neural network weights from the existing nvngx_dlssnr.dll file. The implementation uses a Vulkan layer that intercepts frames and passes them to a separate process for execution on the XMX units. This method allows the technology to function with Windows games running via Proton, DXVK, or VKD3D-Proton. Testing was conducted with titles including Tekken 7, Dead or Alive 5 Last Round, and Mortal Kombat 1.
Performance: From 2 to 14 FPS
While the compatibility is impressive, the performance on the Lunar-Lake architecture is extremely limited. At Full HD (1920 × 1080) resolution with a 0.55 render scale, the frame time is approximately 412 ms, resulting in a mere 2.4 FPS. Even at a lower resolution of 1024 × 768, the frame rate only reaches 6 FPS.
Double-digit frame rates are only achievable at very low resolutions:
- 640 × 360: between 12.5 and 13.5 FPS;
- 512 × 288: nearly 14 FPS.
In actual gameplay scenarios, such as Tekken 7, performance drops further to roughly 10.5 FPS at 640 × 360. The bottleneck stems from the fact that several processing steps still require the full output resolution, placing a heavy load on the bandwidth and compute capabilities of the integrated solution.
It is important to note that these tests were performed on an energy-efficient iGPU. While discrete Arc graphics (such as the B580) are technically capable of supporting these functions, their performance in this specific scenario has not yet been measured. Nevertheless, the project proves that Neural Rendering can fundamentally operate outside of NVIDIA's proprietary ecosystem, following the path previously cleared for Radeon hardware.
Гика
