
In response to a question about larger context windows for LLMs: “yes, on GPUs with suitable memory footprint. Note that since M1 Max, apple GPUs have had crazy 64 - 192 GB of VRAM available for both training and inference workloads. the tests for training our own gen ai tools for artists need 64 GB VRAM, even if slower, it at least works on Apple silicon. Even highest end ada rtx 6000 tops out at 48 GB, so 80 GB A100 or H100 is only nvidia option and those aren’t desktop class cards we’d find en mass out there” @JulesUrbach 28.07.23 $RENDER
