Yeah, super helpful for testing out things

I have to be honest here - there are some smoke and mirrors going on. In this video, I am remotely using a PC with a GPU over the local network as the iterations per second (it/s) is incredibly faster than the mac for SD. For reference i have a M1 Pro and the it/s is about .75 it/s. The PC has an RTX 3060 that outputs roughly 7 it/s. The increased speed and reduced heat and stress on the mac makes it well worth it. Here is a sheet i found really helpful that is worth checking out: docs.getgrist.com/3mjouqRSdkBY/sdperformance It has user reported benchmarks for SD. While no M3 Max silicon is on the list, but a M2 Max gets about 3 it/s (M1 Max was ~2.5) - so i suspect a generational leap in processors probably isn't going to be above 4 it/s. I am a little confused at the technical aspects of the limitations of the mac. From my understanding text gen models like mixtral8x7b can be fully loaded into memory and perform quite well (www.reddit.com/r/LocalLLaMA/comments/18fyn1k/just_installed_a_recent_llamacpp_branch_and_the/) whereas stable diffusion models, which are roughly 4gb can be fully loaded into memory but not perform nearly as well as their GPU counterparts. I believe there is a big difference when it comes to bandwidth of PCI and how the VRAM is used in GPUs (which the macs don't have) vs. regular memory (which the macs do have), but I get lost here as to why this is the case.