Linus Tech Tips Discussion

l controls. The modders report seven available styles, although the current implementation is only affecting character models in this test. [...] There are several reasons not to treat this as a finished DLSS 5 mod. The implementation has not yet been added to a public RenoDX release. HDR is also not working correctly, with current testing focused on SDR while the developers investigate how the new model expects its image data. (VideoCardz)  Quotes Quote At 158.2 MB, this is the largest DLSS DLL NVIDIA has shipped to date—by a comfortable margin. The obvious question is whether that size is real or just development baggage: debug symbols, an unstripped build, that sort of thing. It is not. Breaking the file down by section makes the composition clear: - Executable CPU code: 0.7 MB - Read-only data: 0.2 MB - Compiled GPU kernels and tables: 16.4 MB - Trained model weights: 140.9 MB Roughly 90% of the file is the neural network itself, stored as a single resource blob. The actual program logic is under a megabyte.   [...] Byte count only becomes a parameter count if you know the precision, and that's where the GPU kernels from the previous list come in. 111 of the DLL's 231 CUDA kernels have "_fp8" in their naming, which is a strong hint. [...] So by our own analysis the model is roughly 148 million parameters, and it's running in FP8. For VRAM usage that means, you can expect to see it grow by around that amount—negligible on any modern GPU. On top of that you need some buffers, maybe several hundred MB more, depending on resolution and scale factors, but certainly not gigabytes.   [...] Inside the DLL there are "cubins," one set per architecture. We enumerated every one: 15 cubins containing 231 kernels, and each one targets "sm_120," which is GeForce RTX 50, Blackwell. There is no "sm_75" (Turing), no "sm_80" or "sm_86" (Ampere), no "sm_89" (Ada). Missing cubins alone would not settle it, because CUDA code can also ship an intermediate form (PTX), but we couldn't find this in the DLL at all. The binary also carries an explicit runtime gate reporting an unsupported GPU architecture alongside the minimum required one.   [...] There is no profile table, no game whitelist, no embedded game names, and no engine detection. But here, too, you have to consider that this doesn't look like the final version, and NVIDIA has tuning mechanisms that live outside the DLL, in the driver profile database and NVIDIA App overrides. I suspect that DLSS 5 tuning actually happens on the developer-side, per scene, through masks and sliders passed at runtime.   [...] We found an error string left in the file, which reports a missing module, explains that the block in question is part of the skeleton but that no wrapper has been ported yet, and then it points at a file for the remaining to-do list: ".cursor/rules/crazy-cuckoo.mdc". Tw