Harden Up: Can We Break Your Password With Our GPUs?

Password Cracking: GPGPU-Style

The goal of a brute-force attack is to try multiple passwords in rapid succession. But modern CPUs aren't particularly well-optimized for this. While my Sandy Bridge-based workstation can process about 28 million passwords per second, it still isn't using all of its available CPU cycles. Remember that you're only guessing and checking. The above CPU utilization screenshot shows that, while clock rate is of course going to help, this application is able to take advantage of parallelism, suggesting that even more cores would help accelerate the process. 

Well, as we know, when it comes to parallelized tasks, what a CPU can sometimes do well, the ALUs on a graphics card can often do better. AMD and Nvidia had the video transcoding community convinced of this up until Intel proved a little dedicated logic on its Sandy Bridge architecture could vastly outperform a massively parallel graphics processor. But now it looks like we have another potential application for the general-purpose computing capabilities of today's GPUs.

In the past, GPGPU-based password cracking was limited to academia, where graduate students slaved away over custom code that never saw commercial implementation. That's no longer the case. Now there are two GPGPU tools available to anyone with a credit card: Parallel Password Recovery and Accent Password Recovery.

Parallel Password Recovery is solely optimized for Nvidia's CUDA parallel computing architecture. That's not to say CUDA is better than AMD's Stream initiative. However, Nvidia should be given its due credit. Nvidia's foray into general-purpose GPU computing is a much older one, and the company has given developers access to a plethora of the low-level libraries needed to explore the technology. The response from AMD has been much slower. In fact, we still see issues with getting Stream acceleration enabled in popular transcoding apps. This partially explains why the other solution, Accent Password Recovery, still seems to favor Nvidia's design. Though it recognizes CUDA and Stream, only Nvidia hardware is optimized to attack Zip 2.0 encryption.

And boy is there a boost in Zip 2.0 encryption! With a GeForce GTX 460, we're already pushing over 500 million passwords per second. To give you an idea of what that means, we can now crunch every possible combination of ASCII characters from a password length between one and seven characters in less than 48 hours. Moving up to an eight-character password takes about 168 days.

Of course, Zip 2.0 is rather outdated. WinZip only maintains this scheme for compatibility reasons. AES is the new "it-girl" everyone is fawning over. This scheme is much harder to parallelize, though software developers are certainly trying.

Performance is absolutely hammered once we start trying to recover a password encrypted using the AES cipher. Literally, the mouse cursor on our test system stutters along. Brute-force attacks on AES are clearly slow. To try every possible combination of ASCII characters from a password length of 1 to 7 would take over 13 years.

General-purpose GPU computing is all about parallelism, so if one card's 480 CUDA cores are good, the combined 960 cores from two cards must be better, right?

When we slap two GeForce GTX 570s together and enable SLI, Zip 2.0 encryption starts to look like Play-Doh. Thanks to optimized code, we can push 1.5 billion passwords per second. This is a bit insane. Now we've cut the search time for a one- to eight-character password using all ASCII characters down to almost two months.

Meanwhile, AES security still looks pretty good. If the password is beyond seven characters long, we'll have to spend nearly five years crunching numbers at 500 000 passwords per second.