2 ms·

I tried it, with NVlink. The speedup during inference is negligible. You'll probably benefit more during training.