The point is that a salt would essentially give you a constant speed of cracking versus a speed that scaled with the size of the dataset you have. For example,

The point is that a salt would essentially give you a constant speed of cracking versus a speed that scaled with the size of the dataset you have. For example, the guy in the article cracked 10,000-ish passwords in 16 minutes - 600 passwords per minute. If he'd had a dataset of 1.6 million, he would have cracked over a million (assuming the passwords were of the same quality and he has some good way of storing them on his hardware to allow fast lookup), at a rate of 60,000 passwords per minute, all on his single GPU. However, with a salt he'd be limited to about 1 password per every 27 minutes. And this rate would not increase as the dataset grew larger, which means that even if he had 10x the power as you suggest he could only crack about 20 passwords per hour (more or less). Thus to crack even 1% of a 1.6 million password dataset (16,000 passwords) it would take over a month of non-stop computation on 10 GPU's, versus a single GPU computing 62% (again over a million) in 16 minutes. So even if a compromised website took a month to realize the leak, it would be able to inform it's users in time such that 99% of them would be able to change passwords and avoid their accounts being compromised in time. With no salt, no such luck since as we can see here up to 90% of the dataset would be cracked within a day.