Recent content by OlegSH

You are right. The PS5 Pro's GPU has a 3x3 convolution instruction that results in the 300 TOPS figure. It is quite different to the WMMA. WMMA is more flexible, whereas the convolution instruction is designed for the convolution ops in CNNs. I suppose it performs convs very quickly, maybe even...

https://www.inceptionlabs.ai/news Cool stuff! Now, with diffusion replacing autoregressive word prediction, one can generate tokens 10X faster for a single batch request, with fewer hallucinations and improved metrics as a bonus. Another win for more mathematically dense algorithms...

Both are lossy when quantization is in use, but this is adjustable, allowing the losses to be tweaked per content. In fact, both are quite similar in what they do and what can be achieved with them. The main difference is that DGF is optimized for hardware decoding, while Nanite is optimized for...

As I already said, transcoding from UE5's storage format has never been an issue. You won’t magically achieve a major speedup by accelerating a nonexistent portion of rendering time. Broad compatibility is far more important in that case. And why exactly would DGF somehow remove the necessity...

And why exactly would you want to replace something that works on practically all hardware with something that works on none? The only possible reason could be achieving higher compression ratios for the DFG or perhaps noticeably better performance. However, the claimed 4 bytes per triangle...

It's simpler than that. Typically designers do this to pure mirrors to hide the very obvious cubemap and SSR artifacts. Without the bumpy or scratched mirror surfaces, you would notice the ugly, fat cubemap texels and SSR artifacts in no time.

Besides, AMD's BVH has always been 2 to 3 times the size of NVIDIA's BVH since Turing. Now, with Blackwell's claimed 75% BVH size on top, this would translate into 2.7 - 4x compression ratios for BVH when compared to the current AMD products, which should exceed or match the DGF compression...

Agree, the default sharpening value which was likely finetuned for the CNN model is not a good match for the new TN model.

You did it right. Congradulations!)

DLSS and Native TAA can be swapped in position for each pair.

I would argue they have demonstrated the ability to distill ChatGPT models, using its answers for training and nothing more. The R1 preview even displayed OpenAI’s prices for API calls and answered that it was created by OpenAI) There’s no other way to improve scores in the hard math...

The difference in motion is indeed dramatic. Below are a few comparisons taken in CP2077 during fast forward motion (hence the minor mistmatch). One screenshot in each pair is a native resolution with TAA, while the other is DLSS Performance (with the default sharpening level for DLSS). Display...

It has no relation to AMD’s format, which is lossy, whereas the HW compression for MG clusters in Blackwell is lossless.

What kind of subdivs with displacement mapping are used in this demo then? :)

It filters textures at sharp angles and does so more accurately than HW AF filtering when compared to the ground truth (the 1024 samples rendering) because it filters after the shading rather than before it. It makes sense given that HW AF typically samples an ellipsoidal texture area while the...