I know of one modest demonstration (Feb 2012, link) that claimed the power to make 400 billion guesses a second on a 25 GPU system. In that case, an 8 digit password would be blown in less than 6 hours; sooner depending on the brute-force method. But that assumes the attacker has access to the file that stores the encrypted password. And frankly, that is easy to do, if you have access to the computer itself. Even if you can't get to the HDD, the attacker would simply replace the keyboard with a computer that would send 'keystrokes' much faster than you could type. It might take longer, due to the speed of the USB connection, but human typing rate is not a good reference on this matter.
As a side note.....
On the issue of characters used in a password, this is not quite as simple as most people state. What matters most is what the attacker expects to have to tried, not what characters you chose. In other words, what matters most is what characters EVERYONE in the system uses, not just you. For example, a random sequence of 'X', 'Y' and 'Z' is just as hard to guess as a random sequence of all letters of the alphabet...as long as the attackers doesn't know you prefer X, Y, and Z. But if, despite the availability of 100 digits, it is known to the attacker that everyone is using only X, Y and Z, then the attacker can narrow down the brute-force attack and negate the benefit of 100 digit security system. The principal of this is identical to that of the dictionary attack. This is why sysadmins might force everyone to use different character types; to make sure that a would-be intruder has to try all permutations.
This is not to say the specific characters used in a password don't affect the speed at which it is broken. That is, when someone says "an 8 digit passwords take 10 years break," that 10 years is the MAXIMUM time required. A more accurate statement would be, "it takes 10 years to test all combination of 8 digit passwords." But the fact is that some passwords would be guessed much faster depending on the character selection and attack method. For example, if your password 100-character alphanumeric system (e.g. 0-9......A-Z), and the brute-force attack uses sequential guesses, then a password starting with a '0' will be broken at least 100x faster than a password that starts with LAST character in that sequence (let's call it 'Z'). But this is tricky to deal with since you can never know what order the attacker may use. For example, does the attacker consider A or 0 the first digit? And is Z or 9 the last digit? Or if the attacker knows that everyone uses passwords that starts with characters towards the end of the alphabet, then he/she may try brute-force in reverse-sequence, and the password that starts with '0' will be safer.
Unfortunately, the speed at which passwords are broken is as much about the number of digits as it is the predictability of human behaviour.