Info Facebook unveiled "Neural Supersampling", alternative to DLSS

[Krteq]

Facebook unveiled "Neural Supersampling", alternative to DLSS with no need of specialized HW

Like DLSS it's using motion vectors and trained neural network

Blogpost - Facebook Research - Introducing neural supersampling for real-time rendering
Research paper - Facebook Research - Neural Supersampling for Real-time Rendering

 

[Qwertilot]

Nice NN's will (already) do this sort of thing very well in time.

There's no indication that you wouldn't need specialised hardware though? You would still need to run the Neural Net locally, and that's something that needs acceleration if you want to do it well vs just rendering at full resolution directly.

 

[Bouowmx]

After training, the network models are optimized with Nvidia TensorRT [2018] at 16-bit precision and tested on a Titan V GPU.

Still uses NVIDIA Tensor cores like DLSS to run a neural network with fast matrix arithmetic.

To upscale 4x4 to target 1920x1080, the runtime is 24 ms (standard) or 18 ms (fast). It's not for 60 fps.
In Metro Exodus, NVIDIA says the scaling part needs be fast, only a minor portion of time per frame.

Spoiler: Frame time

However, quality in metrics PSNR and SSIM is better than DLSS. It's a speed-quality trade-off. DLSS was tested in Islands of Nyne (a dead battle royale game?).

 

[antihelten]

Brief overview of the work discussed in this video

Still uses NVIDIA Tensor cores like DLSS to run a neural network with fast matrix arithmetic.

Using TensorRT doesn't necessarily mean that they are using Tensor cores. That's not to say that they aren't, but we cant determine it simply based on that single line from the paper.

Either way though they also state the following:
"In contrast (to DLSS), our work provides full details and evaluation of the method and can be employed without a need for proprietary hardware or software."

 

[PhoBoChai]

DLSS 1 to 1.9 didn't leverage tensor cores at all. It's not until 2.0 version that tensor cores are used, and NV doesn't reveal what its used for exactly.

 

[Mopetar]

Not needing proprietary hardware doesn't mean that it doesn't need specialized hardware or that having the specialized hardware wouldn't allow it to run much faster or more efficiently.

You don't need FP hardware to perform FP calculations, but having them is going to be a lot more effective than trying to produce the same results using integer hardware.