2 ms·

The article doesn't mention this, but I'm curious if this chip accelerates training or nn forward computation?