DLSS 5 Runs in Your Browser Now, But It's Slow as Hell

A developer got NVIDIA's DLSS 5 neural rendering running in a browser via WebGPU, even on macOS. The catch? Each frame takes 1-2 seconds to render.

Imagine firing up DLSS 5 without an RTX 50-series card, without NVIDIA's blessing, and without even leaving your browser. Sounds wild, right? A developer going by the handle MAAN actually pulled it off, porting NVIDIA's neural rendering tech into a plain old web browser using WebGPU. And yeah, it even runs on macOS.

Here's the thing: this completely sidesteps NVIDIA's official pipeline. Right now, DLSS 5 is only officially supported on RTX 50-series GPUs, and you have to call it through the NGX interface inside native apps or the Streamline SDK. NVIDIA's docs never mention WebGL or WebGPU support at all. MAAN's demo does things the unofficial way, pulling model weights from leaked DLSS 5 library files and reimplementing the whole neural network as WebGPU compute shaders. The model weights weigh in at around 147MB, while the JavaScript runtime compresses down to about 1MB.

The demo itself is hosted on Cloudflare Workers and comes with several built-in 3D scenes, adjustable parameters, and a before-and-after comparison view. You can even upload your own 3D models to put it through its paces. It's a neat little playground for anyone curious about how neural rendering actually looks in motion.

But let's talk performance, because this is where things get rough. Testing shows that on an RTX 4090, just loading the neural rendering model takes one to two seconds. Every single frame? Another one to two seconds. For context, hitting 60fps in a game means you've got 16.7 milliseconds per frame. We're not even in the same galaxy here. Real-time gaming is completely off the table.

MAAN says the source code will be published on GitHub. So while this is nowhere near practical for playing actual games, the demo proves something genuinely interesting: DLSS 5's neural network can be decoupled and run on non-NVIDIA hardware. That opens up some potential use cases where real-time performance isn't the priority, like architectural visualization or previewing 3D models. It also hints at a future where you could get AI-powered image enhancement on devices without a dedicated GPU at all.

It's early, it's slow, and it's definitely not going to replace your graphics card anytime soon. But as a proof of concept, it's a pretty fascinating peek at what might be possible down the road. Would you ever mess around with something like this just to see it in action, or is the 1-2 second per frame deal-breaker enough to keep you away? Let us know what you think.