Building a pyrit PC - brief analysis of graphics cards

[clevers]

I've been doing a bit of digging around on the subject of building a pyrit-centric workstation and thought I'd share the analysis here.

If you were looking to buy or build a workstation for pyrit cracking then the benchmarks on the pyrit googlecode page are an interesting reading. The fastest single card there by some margin is the ATI Radeon 5870 with 31k PMK/s. There is good reason to believe that the 5970 would go twice as fast if the current open issues regarding the use of both GPUs are resolved. When you factor in the price difference between the ATI and Nvidia cards it appears to be a slamdunk, ATI are both faster and cheaper. (This is ignoring the temptation to use multiple cards.)

Unfortunately it is not that simple. Whilst OpenCL is touted as the future standard it is the Nvidia CUDA standard that seems to have most traction right now. This is probably not an issue if you just want to use pyrit as it supports both but many other apps are CUDA only. Then there is the issue with drivers. ATI seem to have reputation for issuing broken drivers and the latest 10.2 release doesn't seem to have helped this reputation.

The dilemma gets worse when you consider that the next gen Nvidia cards are due for release next month (delayed again!). The leaked specs and reports suggest slight improvement in graphic performance over the ATI 5870 and that might translate to increased pyrit performance. That said it is likely to come at at least a £100 premium over the 5870.

So it seems the choices today are to put up with ATI and go with a 5870 (which can be got in the UK for < £300) or wait for the new Nvidia GTX 480. Either way I've decided to hold off and see what the launch of the new Nvidias do to card prices across the board and also to see a real pyrit benchmark on one of those cards.

[purehate]

Also just to add, the claim of 31k pmk.s was made by elcomsoft and not by pyrit. This is a profit based company who grossly overexerted the figures for ntlm cracking as well. They claimed 1.3 billion ntlm per second on a 295 gtx when if fact it is more like 600 million.
ATI has also always had the worst linux support as well, whereas nvidia has a entire section working on CUDA/linux type stuff at all times. Not to mention none of the other cool GPU tools work with ATI. I will stick with the team that is in active linux development, even if it mean getting a few less pmk/s

Just my 2 cents

If you are interested here is a breakdown of my latest CUDA/Nvidia box
http://tools.question-defense.com/20...a-fat-upgrade/

[clevers]

Are you sure it's an Elcomsoft number? Just looking on their site and they do have some numbers but not the 5870. They are claiming 103,000 PMK/s for the AIT 5970 and have the GTX 295 at 22,000 (which is probably 10% higher than pyrit).

That box you have built is a monster, I was aiming for something a little more modest for some OpenCL / CUDA messing about and a bit of pyrit PoC type stuff. Maybe i5 based < £1k type setup.

[kidFromBigD]

FTA: "No bad for a couple of Kentucky Boys"

--> Dude, that server rocks! Love the Cooler Master V8 sticking out. Good work.

[kafteras]

Well CUDA is more wide accepted but for pyrit ATI is the answer with 5870 or 5970.

Issue 140 - pyrit - CAL++ version of pyrit - Project Hosting on Google Code

They are going for new code and hopping for 55k-60k with 5870. We'll see.

[snowmanwithahat]

Yeh.... from what I've seen ATI is doing alot better.

the claim of 31,000 though was through the use of OpenCL and not the pyrit-stream

I'd like to see nvidia cards benchmarked with open-cl as well and not just the CUDA implementation of pyrit. Also for those of you interested.... ATI drivers seem to be bugged and won't let you use multiple cores with pyrit. The 2nd core of a 5970 for example produces errors beyond redemption.

That kind of power from a single 5870 is very very enticing though. I might consider selling my GTX 285 and 9800 GTX if I can pick up a 5870

After seeing the claims of 31,000 I just re-ran the benchmark to see what mine was pushing out.... If the claims that a 5870 can in fact push out those kind of numbers it might be worth it.

root@bt:~/cse422# pyrit benchmark
Pyrit 0.3.1-dev (svn r228) (C) 2008-2010 Lukas Lueg pyrit - Project Hosting on Google Code
This code is distributed under the GNU General Public License v3+

Running benchmark (23053.3 PMKs/s)... /

Computed 24288.70 PMKs/s total.
#1: 'CUDA-Device #1 'GeForce GTX 285'': 14037.1 PMKs/s (RTT 2.7)
#2: 'CUDA-Device #2 'GeForce 9800 GTX/9800 GTX+'': 6885.3 PMKs/s (RTT 2.9)
#3: 'CPU-Core (SSE2)': 788.6 PMKs/s (RTT 3.1)
#4: 'CPU-Core (SSE2)': 787.9 PMKs/s (RTT 3.0)

After picking up this 9800 GTX I figured it'd help quite a bit.... it's been pretty ineffective though :/

[azboy]

Don't mean to bump an old thread but was wondering if OP ever went thru with building the pyrit box using the ATI card? I'm looking to do the same but there's very little discussion on BT forums about using the HD 5970.

I think this is just a single HD 5970 card getting over 100,000 PMKs someone posted on the pyrit site.

pyrit benchmark
Pyrit 0.3.1-dev (C) 2008-2010 Lukas Lueg pyrit - Project Hosting on Google Code
This code is distributed under the GNU General Public License v3+

Running benchmark (95467.3 PMKs/s)... -

Computed 100756.52 PMKs/s total.
#1: 'CAL++ Device #1 'ATI CYPRESS'': 49512.9 PMKs/s (RTT 2.8)
#2: 'CAL++ Device #2 'ATI CYPRESS'': 49408.2 PMKs/s (RTT 2.7)

To compare pureh@ate's 4 GTX 295 is getting around 40,000 PMKs/s. Imagine if you had 4 HD 5970s!

[purehate]

Just to clarify, that screen shot of mine you have posted is only 2 295-gtx. Each card shows up as 2 cards in pyrit. I do have 4 but I get about 100,000 pmk/second.

On the subject of ATI cards, I do not have one since I use cuda for lots of other stuff but I have heard that pyrit miss's lots of easy passwords while using ati gpu suport.

[TheLaw]

"On the subject of ATI cards, I do not have one since I use cuda for lots of other stuff but I have heard that pyrit miss's lots of easy passwords while using ati gpu suport."

Interesting that. My setup atm is a modest gtx275 - approx 17000 pmks/s without x running. I have found that in some test cases I ran using cuda and pyrit, that the password was not found. However I would run the same test the following day and it would be found. I also get segmentation errors 1 out of every 5 passthrough attempts.
Due to these factors my confidence in the above mentioned programs is a little shaky.

[hhmatt]

"On the subject of ATI cards, I do not have one since I use cuda for lots of other stuff but I have heard that pyrit miss's lots of easy passwords while using ati gpu suport."

Interesting that. My setup atm is a modest gtx275 - approx 17000 pmks/s without x running. I have found that in some test cases I ran using cuda and pyrit, that the password was not found. However I would run the same test the following day and it would be found. I also get segmentation errors 1 out of every 5 passthrough attempts.
Due to these factors my confidence in the above mentioned programs is a little shaky.

I don't have a graphics card powerful enough to get any advantage using pyrit. But I do know that with password cracking there is such a thing as too fast. Especially when we are talking about doing this across the network. You will have to find the right balance of speed and reliability. Pyrit as a program can only interpret the results soo fast and the program you are attempting to break can only accept and reply to your attempts soo fast. There is a lot of room for error outside of pyrit's control.

Have you documented and reported your segmentation faults?