A new GitHub project, dlss-nr-on-intel, has successfully implemented DLSS 5 neural rendering on an Intel Arc 140V GPU. Developed by AI agents Claude Opus 5 and Astra, the Apache 2.0 licensed code uses Vulkan shaders and Intel XMX units instead of Nvidia code.
The implementation utilizes a symmetric U-Net with 71 blocks and five Swin stages, requiring users to extract 73,841,889 FP16 parameters from their own nvngx_dlssnr.dll. It runs on Linux via Mesa ANV, using VK_KHR_cooperative_matrix for matrix operations.
Performance remains experimental, reaching 10.1 Fps at 512x288 but dropping to 1.2 Fps at 1080p. With a base load of 17ms plus 488ms per megapixel, this proof-of-concept shows neural rendering potential on non-Nvidia hardware despite its current speed limits.