TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
OpenDLSS is a public Vulkan reimplementation of NVIDIA’s DLSS 5 Neural Rendering network, with its developers reporting byte-for-byte matches at 75 block boundaries against reference captures. It requires users to supply the model weights and currently targets Windows PCs with NVIDIA Ada-generation or newer GPUs; the project is not DLSS Super Resolution.
OpenDLSS, a project published on GitHub, offers a Vulkan reimplementation of NVIDIA’s DLSS 5 Neural Rendering network and says its output matches reference captures byte for byte at 75 network block boundaries. The release matters because it makes a technically detailed implementation available outside NVIDIA’s software stack, but users must provide the model weights, and the project requires a recent NVIDIA GPU and specific driver support.
The project describes the network as the same 71-block Swin/ViT model as DLSS-NR build 310.8.0. It uses six pooling levels, FP8 E4M3 activations with FP16 accumulation, and a reported 141 MiB of weights. OpenDLSS says matching is not limited to the final image: its intermediate results agree with the reference at all 75 block boundaries. That is a claim made by the project; the supplied materials do not include an independent evaluation.
OpenDLSS characterizes the model as a generative neural rendering network, not an upscaler. It processes a frame at the same resolution, using a low-dynamic-range proxy, three Gaussian-noise lanes, a reprojected previous-frame output and five conditioning scalars. The network returns four values per pixel: an RGB residual and a temporal-blend logit. The project says the demo implements that temporal feedback path, while its command-line tool processes single frames without history.
The repository also includes a separate browser WebGPU implementation, which its developers say produces the same bytes without tensor cores or FP8. On the project’s stated comparison, it takes 72 milliseconds at 512-by-512, against 2.7 milliseconds for the Vulkan implementation at that resolution. The project’s benchmark table reports Vulkan runtimes on an RTX 4070 SUPER of 2.8 milliseconds at 768-by-768, 7.8 milliseconds at 1920-by-1080, 12.6 milliseconds at 2560-by-1440 and 29.3 milliseconds at 3840-by-2160. Those are project-reported minimums over 40 frames, not independent benchmark results.
A Reimplementation Outside NVIDIA’s Stack
OpenDLSS gives graphics developers and researchers a way to inspect and run an implementation of a complex neural rendering pipeline using Vulkan, rather than relying solely on NVIDIA’s own implementation. Its reported intermediate-layer matches may also help technical users examine how the network’s computations behave, provided they can obtain the required weights and reference material.
The release does not mean that DLSS 5 is now broadly available across hardware or that NVIDIA has released its model parameters. OpenDLSS lists requirements including Windows, an NVIDIA Ada-or-newer GPU, and drivers exposing several specified Vulkan and NVIDIA extensions. Its performance figures are tied to one listed graphics card and project test method. The practical reach is consequently narrower than the term “Vulkan reimplementation” alone might suggest.
The distinction from DLSS Super Resolution matters, too. The repository explicitly says DLSS-SR is a different network and is not implemented. OpenDLSS concerns neural rendering that modifies a frame at its existing resolution; it should not be described as a general replacement for NVIDIA’s upscaling technology.
NVIDIA DLSS 5 neural rendering GPU
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Project Implements
The repository identifies NVIDIA’s DLSS 5: Generative Neural Rendering report as its model reference. Its implementation combines C++ host code, Vulkan shader kernels and generated PTX kernels. The project describes the Vulkan route as using cooperative-matrix FP8 operations and fused network operations; its faster route uses NVIDIA-specific PTX features. A reference route and tools for parity checking are also included.
Running the network requires users to supply a model directory in the format described by the project. The source does not say that OpenDLSS distributes the model weights. The optional interactive demo uses Filament and glTF scenes, while the core command-line tool can benchmark or compare runs against fixtures. The source lists supported hardware and software requirements, but does not provide evidence that the implementation has been validated on other operating systems or GPU vendors.
“A Vulkan reimplementation of NVIDIA’s DLSS 5 Neural Rendering network, bit-exact against the original.”
— OpenDLSS project description on GitHub
As an affiliate, we earn on qualifying purchases.
Weights and Independent Verification
The supplied project materials do not establish whether model weights are publicly available or how users can obtain them. The project says users must provide a model directory, leaving access to the parameters as a practical dependency. Its claims of bit-exact output and the listed performance measurements are presented by the project; no independent test results or external confirmation are included in the source material.
It is also unclear how the implementation performs across different supported Ada-or-newer GPUs, driver versions or scenes beyond the examples and fixtures described by the repository. The benchmark table gives minimum times over 40 frames and notes GPU clock changes under sustained load, but does not provide an independent test setup. The source does not specify a release date, licensing terms for every project component, or any response from NVIDIA.
As an affiliate, we earn on qualifying purchases.
Testing the Released Implementation
The next step for readers with compatible hardware is to inspect the repository’s model-format and build instructions, then test the implementation using weights and fixtures they are authorized to use. OpenDLSS provides parity and verification commands for comparing results, along with benchmark and profiling tools. The project’s documentation also describes its frame-history path and demo setup.
Further confirmation would require independent reproductions of the reported byte-level matches and performance figures, plus clarity on model-weight access and distribution. Until those details are established, the project is best understood as a technically specific implementation and a set of developer claims, not as evidence that NVIDIA’s complete DLSS software or model has become generally available.
NVIDIA Ada generation graphics card
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is OpenDLSS?
It is a GitHub project that implements NVIDIA’s DLSS 5 Neural Rendering network using Vulkan, with a separate browser-based WebGPU implementation.
Does OpenDLSS include the model weights?
The supplied project description says users must provide a model directory. It does not establish that the weights are distributed with the project or explain how to obtain them.
Is this a DLSS upscaler?
No. The project says the network processes and outputs frames at the same resolution, and explicitly identifies DLSS Super Resolution as a different network that is not implemented.
What hardware does it require?
The repository lists Windows, an NVIDIA Ada-or-newer GPU and a driver exposing several specified Vulkan and NVIDIA extensions. It does not document broad support for other GPU vendors.
Have the exact-match claims been independently confirmed?
Not in the supplied source material. OpenDLSS reports byte-for-byte agreement at 75 block boundaries, but the materials provide no independent verification.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
