3 ms·

True, but why not just run it on the host of the GPUs then? I guess the advantage would be on-demand GPUs.