| ▲ | KeplerBoy 4 days ago | |
Also it's not just about running an obviously worse quant. Running different GPU kernels / inference engines also matters. It's easy to write an implementation that is faster and thus cheaper but numerically much noisier / less accurate. | ||