TL;DR
Get the latest gadgets delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
OpenDLSS is a public project that says it reproduces NVIDIA’s DLSS 5 neural rendering network in Vulkan, matching reference outputs at all 75 block boundaries. It also offers a separate WebGPU implementation, but users must provide the model weights, and the Vulkan build requires a recent NVIDIA GPU and specific driver features.
The GitHub project OpenDLSS says it has reimplemented NVIDIA’s DLSS 5 neural rendering network in Vulkan, with outputs matching its reference at all 75 block boundaries. The release also includes an independent browser-based WebGPU port, but it does not provide the model weights: users must supply a compatible model directory.
According to the project documentation, its Vulkan version runs the same 71-block Swin/ViT network as DLSS-NR build 310.8.0, using E4M3 FP8 activations with FP16 accumulation. OpenDLSS says the intermediate data matches byte for byte across all 75 block boundaries, not only in the final output. Those are claims made by the project; the supplied material does not include an independent verification.
The network takes a rendered frame alongside other inputs, including three lanes of Gaussian noise, a reprojected prior output and five conditioning values. It returns an RGB residual and a temporal-blend logit for each pixel. The project describes this as neural rendering rather than upscaling: input and output use the same resolution, and the network re-renders a frame produced by the game engine.
OpenDLSS reports that on an RTX 4070 SUPER, its Vulkan implementation took 7.8 milliseconds per frame at 1920×1080 and 29.3 milliseconds at 3840×2160, measured as the minimum across 40 frames. The figures cover the whole network, which runs through 241 dispatches. The project says the GPU alternated between two clock states during sustained load, making median times a few percent higher; these are project-reported measurements, not an independent benchmark.
A New Route to Testing Neural Rendering
OpenDLSS gives developers and researchers a way to examine and run a claimed reproduction of a neural rendering network outside NVIDIA’s own implementation. The project includes its network graph, Vulkan kernels, a CPU arithmetic reference, tools for comparing outputs and a demo built around the Filament renderer. This could make it easier to study how the network processes frames and to test its behavior in a separate rendering environment.
The browser port broadens access to the implementation, but it is not equivalent in speed or hardware path. OpenDLSS says it runs without tensor cores, FP8 or cross-block fusion and measured 72 milliseconds at 512×512, compared with 2.7 milliseconds for the Vulkan route at that resolution. That result is a project-provided comparison, and it should not be read as a general benchmark across browsers or systems.
The project does not reproduce all DLSS features. Its documentation says DLSS Super Resolution is not implemented, because it is a different network. The release is therefore relevant as a specific reimplementation of the neural-rendering component, not as a replacement for the broader DLSS suite.
NVIDIA RTX 4070 SUPER graphics card
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What OpenDLSS Implements
NVIDIA describes DLSS 5 as generative neural rendering, according to the project’s documentation, which points readers to NVIDIA’s report on the system. OpenDLSS characterizes the model as a U-net combining shifted-window transformer blocks with a global vision transformer at its deepest level. It says the network has six pooling levels and 141 MiB of weights.
The project separates its Vulkan reference route from faster, hardware-specific kernels and includes a WebGPU port for browser execution. Its demo connects the network to a Filament-based renderer and includes a temporal feedback path. The command-line tool, by contrast, processes single frames without history, which OpenDLSS says is the setup used for its reference captures.
Using the Vulkan version requires Windows, an NVIDIA Ada-generation or newer GPU, and a driver exposing several specified Vulkan and NVIDIA extensions. The project says users need to provide the weights themselves in the documented model-directory format. It explicitly identifies DLSS-SR as out of scope.
“The same bytes in a browser, with no tensor cores and no FP8.”
— OpenDLSS project documentation
As an affiliate, we earn on qualifying purchases.
Independent Verification and Model Access
The material provided does not identify an independent audit of the bit-exactness claim or the performance results. OpenDLSS describes parity and verification tools, but it does not establish that outside researchers have reproduced those results. The documentation also says users must supply the model weights; the source does not clarify how readers can obtain them or whether access is restricted by separate terms.
It is also unclear how the implementation performs across different supported GPUs, drivers or real games. The listed timings are for one RTX 4070 SUPER and specific test resolutions. The project’s demo and test fixtures can help evaluate its behavior, but the supplied report does not provide independent visual-quality comparisons or broad compatibility results.
As an affiliate, we earn on qualifying purchases.
Testing Across GPUs and Scenes
The next practical step is for developers to build the project, provide compatible weights and run its parity and verification tools against available fixtures. OpenDLSS also provides a demo for loading scenes, which could let users examine the network’s output and temporal behavior in a renderer.
Further reporting would be needed to confirm the implementation’s matching results independently, establish how the required weights may be accessed and compare performance across hardware. The project’s documentation does not announce a release schedule or a broader compatibility plan.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is OpenDLSS?
OpenDLSS is a GitHub project that says it reimplements NVIDIA’s DLSS 5 neural rendering network in Vulkan, with a separate browser-based WebGPU version.
Does it include NVIDIA’s model weights?
No. The project documentation says users must supply the weights in a compatible model directory.
Is OpenDLSS an upscaler?
The project says no: its network takes and returns frames at the same resolution. It also says DLSS Super Resolution, a different network, is not implemented.
What hardware does the Vulkan version require?
The documented requirements include Windows, an NVIDIA Ada-generation or newer GPU, and a driver that supports specified Vulkan and NVIDIA extensions.
Have the matching results been independently confirmed?
The source material provides OpenDLSS’s own claim of byte-for-byte matches at 75 block boundaries, but does not cite independent verification.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
