Are you sure its not using tensor cores? Because even previous implementation used tensor cores in some degree.
Thats even in your old post:
https://www.resetera.com/threads/dl...implement-it-well-anyway.168195/post-28857425
Maybe something about reporting in NSight is broken.
Yes, I'm quite sure. I don't think NSight is broken, it's a quite precise tool and if there would be something wrong with it, it wouldn't even let me capture a frame.
Yes, that's why I'm so surprised. DUTTM and WY used Tensors to great extent. I was under the impression that the tensor cores are the main part of how they were able to deliver so much improvement at such high framerates, but looking at Control it seems like it has more to do with more training and fine tuning.
The performance with the new implementation is a tiny bit lower than with the last shader implementation on the same render resolution, so its certainly more demanding. If it was using the tensor cores, it would run faster instead, relieving the shader ALUs.
If Nvidia can be trusted, this is a DL neural network. Could be running in FP32 precision on the shader cores for sure, I don't think such improvements can be made on an image processing software like before.
See the power of machine learning:
720p no MSAA full RT, upscaled to 1440p. This is the most demanding scene in the game so it's really heavily GPU limited, ideal for testing.
720p no MSAA full RT, reconstructed to 1440p using DLSS
As you can see, DLSS only takes 5 FPS for reconstruction and delivers an immensive jump in image quality. It's magic.
It's possible that their deep learning model is so fast and efficient that tensor cores are just not needed anymore, I mean yeah it comes pretty close to upscaling performance. If that's the case, I'd rather want them to use the tensor cores for ML denoising and for DirectML stuff in the future.
