Randomoneh wrote: ↑20 Dec 2025, 15:07
What are GPU/CPU requirements specifically for scanning backlight simulation without RIFE interpolation? Would 1650 Super or 7900XT work?
So, I haven't testing on those GPUs so I have to speculate to some degree based on my testing and understanding of the performance considerations.
The main performance considerations/expectations I talk about in my documentation/requirements is specifically targeted at the RIFE interpolation. Because it is so high quality and demanding, I have built a variety of scaling options to allow you to tune performance to your hardware. Resolution scaling is one setting that I make sure works well and provides a seamless way to handle various content.
With that in mind, BFI and CRT scanout simulation is run through an entirely separate process and does not access any of the convolutional neural network acceleration APIs like TensorRT, OpenVino, or MigX.
The only performance factors revolve around video decoding and more importantly, video data transfer bandwidth. The CRT scanout simulation shader itself is a negligible performance consideration but inserting multiple new frames per source frame requires a lot of copying on the fly and moving texture data around between the CPU and GPU before rendering to the framebuffer.
That is all to say, PCIe bandwidth should be an important factor as well as memory bandwidth both in system RAM and VRAM. Capacity isn't much of a concern, just speed. ROPs may also factor in. I would expect the 1650 Super might be a little more limited with its PCIe 3.0 speeds and small 128bit bus. Using the resolution scaler would limit the bandwidth required and should allow it to run smoothly.
The 7900XT has PCIe 4.0 and a much larger 320bit bus and nearly 70% faster memory clock resulting in more than 4x the bandwidth. I would expect the AMD 7900XT to handle BFI or CRT scanout processing quite well.
I'll also add that I have focused heavily on providing the best possible performance and also platform support for the ultra high quality interpolation which is the most demanding process. If there is interest, I can work on refining resource consumption and hopefully make the BFI and CRT scanout run even faster. Make sure to vote in the poll if that's a request.