DLSS 5 Neural Rendering Reimplemented on Intel Arc 140V

An independent open-source project has reimplemented the inference path used by NVIDIA DLSS 5 Neural Rendering so it can execute on Intel Xe2 graphics. The developer built and measured the project on the Arc 140V integrated GPU in a Lunar Lake system under Linux, using Intel’s XMX matrix hardware through the Vulkan VK_KHR_cooperative_matrix extension rather than CUDA or NVIDIA NGX.

This is materially different from making an NVIDIA DLSS DLL ignore a hardware check. The repository describes a resident 71-block U-Net implementation whose matrix operations run through Vulkan cooperative matrices. A Vulkan layer intercepts presented frames and hands them to a daemon that executes the neural network, then composes the generated residual back into the game frame.

The project’s current published measurements show why this should be read as an architecture experiment rather than a gaming feature. On the Arc 140V test machine, its end-to-end path measures about 99 ms per processed frame at a 512×288 swapchain with a 0.35 render scale, about 180 ms at 854×480 with a 0.50 scale, and roughly 850 ms at 1920×1080 with a 0.55 scale. Those figures correspond to about 10.1, 5.6 and 1.2 processed frames per second respectively before treating the implementation as a normal real-time graphics feature.

The repository also makes clear that render scale is not the only cost. Some full-frame work still follows the output extent, so simply reducing the neural network’s internal extent cannot make the complete pipeline scale like a conventional upscaler. The author characterizes the present live mode as a slideshow and provides a separate photo-mode workflow for one-shot processing.

The useful result is portability of the inference workload, not competitive performance. The project says its graph uses FP16 operands with FP32 accumulation on Xe2 because that architecture does not expose the FP8 path used for the large matrices in the source model. It also needs about 2.3 GiB of device-buffer memory at 720p on the integrated GPU, where graphics memory is shared with system RAM.

That makes the experiment a concrete example of how a neural graphics network can be mapped onto another vendor’s matrix hardware when the execution graph and weights are available. It does not establish that Intel Arc can run NVIDIA’s production implementation efficiently, nor that the quality, temporal behavior or performance is equivalent to supported DLSS 5 on GeForce hardware.

The author says AI agents produced the code, measurements and development notes while the human supplied the machine, binary, direction and decisions. The repository documents failed hypotheses and withdrawn conclusions as well as successful measurements, and includes roughly 190 checks in its current test suite. That provenance matters because the results remain project-reported measurements rather than independent GPU benchmarking.

The project has also started accommodating discrete Arc hardware, but its current documentation says discrete-card behavior has not yet been measured to the same standard as the Arc 140V development system. A faster Arc GPU with dedicated memory could change the numbers substantially; until controlled measurements exist, extrapolating desktop Arc performance from the Lunar Lake iGPU would be speculation.

A separate recent community project demonstrated a DLSS 5-style neural workload through WebGPU in a browser, including on non-NVIDIA hardware. This Intel project answers a different question: whether the recovered inference graph can be implemented directly against Intel Xe2 matrix hardware and inserted into an actual Vulkan game path.

Both experiments remain far from replacing NVIDIA’s supported runtime, but together they narrow the hardware dependency of the underlying computation. The next useful evidence would be repeatable measurements on discrete Arc hardware and further independent validation of image quality and temporal behavior—not another claim that Intel GPUs have somehow become officially DLSS-compatible.