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 community Vulkan reimplementation of NVIDIA’s DLSS 5 Neural Rendering network, with the project reporting byte-for-byte agreement at all 75 block boundaries against reference captures. It requires users to provide the model weights and runs on recent NVIDIA GPUs; the project also includes a slower browser-based WebGPU version.
A GitHub project called OpenDLSS has published a Vulkan reimplementation of NVIDIA’s DLSS 5 Neural Rendering network, reporting byte-for-byte matches with reference captures at all 75 network block boundaries. The project is not an NVIDIA release: it requires users to supply the model weights and is designed for compatible recent NVIDIA GPUs.
The project describes the network as a 71-block shifted-window transformer with a global vision transformer layer, arranged across six pooling levels. It uses E4M3 FP8 activations with FP16 accumulation and has 141 MiB of weights. OpenDLSS says its parity checks cover every block boundary, not only the final rendered output; the claim is based on comparison against supplied fixtures.
OpenDLSS characterizes DLSS 5 Neural Rendering as a generative rendering network, rather than an upscaler. It takes an already rendered frame along with noise lanes, a reprojected previous-frame output and five conditioning values, then returns an RGB residual and a temporal-blend logit for each pixel. The project says input and output resolution are the same.
On an RTX 4070 SUPER, the authors report minimum whole-network times of 2.8 milliseconds at 768-by-768 pixels, 7.8 ms at 1920-by-1080, 12.6 ms at 2560-by-1440 and 29.3 ms at 3840-by-2160. Those figures are the minimum across 40 frames, with 241 dispatches at each resolution. The project notes that GPU clock changes under sustained load can make median results a few percent slower.
A Reimplementation Outside NVIDIA’s Stack
OpenDLSS makes a technically detailed rendering network available in a separate implementation, offering developers and researchers a way to inspect its computational structure and test its behavior without using NVIDIA’s original runtime. The reported intermediate-level parity is more specific than a claim that two final images look similar: it gives users a stated basis for checking the implementation against reference data.
The project’s practical reach is limited by its hardware and model requirements. It calls for Windows, an Ada-generation or newer NVIDIA GPU, a driver exposing several listed Vulkan and NVIDIA extensions, and the model weights, which are not included as part of the implementation. The source does not establish that the project can be used with arbitrary games or that it offers a drop-in replacement for NVIDIA’s software.
NVIDIA RTX 4070 SUPER graphics card
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Network Recreates
The project identifies its target as the network in DLSS-NR build 310.8.0, and points readers to NVIDIA’s report, “DLSS 5: Generative Neural Rendering,” for the model description. Its scope is specifically Neural Rendering, which the project says modifies a frame at the same resolution; it explicitly says DLSS-SR is not implemented because that is a different network.
OpenDLSS includes two routes: Vulkan kernels intended for compatible NVIDIA hardware, and a separate browser-based WebGPU implementation. The project reports that the WebGPU version also matches the same captures, but without tensor cores, FP8 or inter-block fusion. Its stated comparison is 72 ms at 512-by-512 for WebGPU versus 2.7 ms for the Vulkan route at that resolution.
“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.
Independent Validation and Model Access
The source material is the OpenDLSS GitHub project; it provides no independent benchmark or third-party confirmation of the bit-exactness and performance claims. The stated parity depends on reference fixtures, but the supplied material does not explain who produced those captures or how they were obtained. Users must also supply a model directory containing the weights, and the source does not establish their availability, licensing terms or redistribution status.
It is also unclear whether the implementation supports use beyond the project’s test fixtures and demo, or how it performs across different compatible GPU models and drivers. The project says its demo implements a temporal feedback path, while its command-line tool processes single frames without history; these modes should not be treated as equivalent.
DLSS 5 neural rendering model weights
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing the Release on Supported GPUs
The next step for readers interested in the claims is to inspect the project’s code, model-format documentation and parity fixtures, then test the build on hardware and drivers meeting its requirements. OpenDLSS provides commands for benchmarking, profiling dispatches and checking parity against a fixture, as well as a demo using the Filament renderer.
No formal independent review, broader hardware support or future release schedule is stated in the source material. Whether the implementation becomes useful beyond evaluation will depend on access to the required weights, reproducible testing and confirmation of results on systems beyond the project’s reported RTX 4070 SUPER setup.
As an affiliate, we earn on qualifying purchases.
Key Questions
Is OpenDLSS an official NVIDIA release?
No. The source identifies it as a GitHub reimplementation of NVIDIA’s DLSS 5 Neural Rendering network, not as software released by NVIDIA.
Does OpenDLSS include the model weights?
The project says users must provide the weights in a model directory. The supplied description does not confirm whether or where those weights can be obtained.
Does it upscale images?
No, according to the project. It describes the network as processing an already rendered frame and producing output at the same resolution; it also says DLSS-SR is not implemented.
What hardware does the Vulkan version require?
The listed requirements include Windows, an NVIDIA Ada GPU or newer, and a driver exposing the specified Vulkan and NVIDIA extensions. The project’s reported performance figures are from an RTX 4070 SUPER.
Has the reported bit-exactness been independently verified?
The supplied source reports byte-for-byte parity against fixtures, including at 75 block boundaries, but provides no independent verification. The claim should be treated as the project’s own report unless separately confirmed.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
