@lordofmetis 我记错了

“From a crosspost of this same question, someone helpfully pointed out that the memory bandwidth benefits of the P100 aren't all that helpful once the model is loaded, at least when inferencing. For training and fine tuning, it matters more. This explains why the P100 is geared towards training and the P40 towards inferencing, and so the larger vram pool is probably the way to go if like me, you're mostly planning to use existing models. Hope this helps.”