DLSS 5 Cracked Wide Open: Now Running on Non-Nvidia Hardware

A developer's open-source Vulkan reimplementation of DLSS 5's neural network runs bit-for-bit identical to Nvidia's version — and it even works in a browser.

Bet you didn't think DLSS 5 could run without an RTX 50-series card. Well, a developer just proved otherwise.

A dev going by MAAN has released OpenDLSS-NR, an open-source Vulkan reimplementation of Nvidia's DLSS 5 neural rendering network. The big deal here? It's the first time DLSS 5 has been shown running on non-RTX 50-series GPUs.

According to the report, OpenDLSS-NR faithfully reproduces all 71 neural network blocks from DLSS 5's runtime DLL version 310.8.0. MAAN claims the output is "bit-for-bit identical" to Nvidia's official implementation — not just the final image, but every single byte across all 75 intermediate block boundaries. That's a level of precision that's honestly kind of wild.

The network itself uses a U-Net architecture, mixing shifted-window Transformers with a global vision Transformer. It runs on FP8 activations and FP16 accumulation precision, with model weights weighing in at roughly 141MiB.

On the performance side, MAAN's benchmarks on an RTX 4070 SUPER are pretty encouraging. At 768×768, execution takes just 2.8 milliseconds. Bump it to 1080p and you're looking at 7.8ms, 1440p hits 12.6ms, and 4K lands at 29.3ms. Those numbers suggest the full DLSS 5 network can already run efficiently on Ada Lovelace architecture.

But the real eyebrow-raiser is the WebGPU port. This version doesn't rely on Nvidia Tensor Cores or native FP8 support at all — it runs entirely in your browser. At 512×512, it takes 72ms. Sure, that's way slower than the native 2.7ms, but it proves something important: DLSS 5's neural network can be decoupled and run on non-Nvidia hardware.

There are still some catches. The repository doesn't include Nvidia's proprietary model weights, so you'll need to supply those yourself. It's also not a plug-and-play game mod — it's a standalone implementation with a demo renderer built on Google's Filament engine. It supports motion vectors and temporal feedback, but doesn't implement DLSS super resolution.

For RTX 40-series owners who can't use official DLSS 5, this project shows there's another path forward. It's still a long way from actual in-game deployment, but it proves DLSS 5's core tech isn't completely tied to the latest hardware.

What do you think — could we eventually see DLSS 5 running on older GPUs in real games? Drop your thoughts below.