![]() ![]() I'm showing single cpu core maxed averaging 95%, second core at below 50% and the rest nearly idle. If you wanna really see how you card works in fp32, feel free to use this from settings -> stable diffusion section Why? Because it is already in use in other parts of automatic's code, its enabled by default for gpu which supports fp16. It has zero effect on 4090, atleast on windows. It was made to force 16xx geforce cards to use fp16. But at least the "if" statement will then do what you want. ![]() But if you are using your 4090 for A1111 then another option would be to buy a (7, 5) based GPU and plug it in. The code tests if any card you have just happens to be a (7, 5) sm_75. Your code doing nothing, because it is already in use in if method, which is commented clear "# enabling benchmark option seems to enable a range of cards to do fp16 when they otherwise can't"Ĭomments don't affect behavior. The bottom line is you are not that far from the theoretical max for your setup. Preferable I'll find some Windows system expert, I'm a Linux expert, to see if there's something like a 10% system overhead stealing CPU time. One day I'll boot the Windows side of my system and look at this myself. The app has DIRECT access to the 4090 without needing to go through the OS. On Linux there is no system time overhead. I'm had one Windows user report system/kernel time overhead within the A1111 process. One where my monitor is plugged into the Intel Integrated GPU leaving my 4090 dedicated for AI computations. One possible issue, which I haven't tested, is that I have a PURE setup. Windows users often struggle to get the same perf as those that are on Linux. Given your cpu is 4.8 GHz vs my 5.8 GHz I would calculate an "estimated" max of 30.28 for your setup. If I don't use the benchmark setting nor -opt-channelslast I get 36.6 it/s with sd1.5. I hope you aren't reporting the first run after starting a1111. If someone has an idea on how to streamline this process feel free to share it, this is just a personal proposal beacuse I haven't seen a centralized data collection of this sort.īeta Was this translation helpful? Give feedback. Post your system info with this extension.Set COMMANDLINE_ARGS= -listen -xformers -autolaunch -port 7861 -enable-insecure-extension-access Post the CLI line Timestamp (we want the final number to measure performance "6.26it/s"):.Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 69, Size: 512x512, Model hash: b2c0fd85d2, Model: StableDiffusion1.5 Negative prompt: monochrome, lineart, blurry, low quality Image of a GPU in flames, (graphics card:1.1) with (double fan:1.2), chrome, realistic, Nvidia RTX, Radeon graphics, studio lighting, product advertisement, Generate the following image with these parameters:.To achieve this I propose a simple standarized test. So the idea is to comment your GPU model and WebUI settings to compare different configurations with other users using the same GPU or different configurations with the same GPU. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |