Potential optimization for older GPU's like the 1080ti

[pflky]

So with GPU's like the 1080ti that have a crippled FP16 performance, FP32 runs faster but consumes more memory. This should be able to be counteracted, by running Stable Diffusion in FP16 memory wise, but spoofed to run on FP32 cores as if it was FP32, thereby gaining performance benefits of FP32 while keeping an FP16 memory footprint. Maybe there is already a setting for this I'm not aware of?