Password recovery : Theoretical and practical results

And finally a summary table for the sake of which everything was started. Comparison rates for different types of passwords. Were used:

• crark http://www.crark.net
• Advanced Archive Password Recovery http://www.elcomsoft.com/archpr.html
• Elcomsoft Wireless Security Auditor http://www.elcomsoft.com/ewsa.html
• Advanced PDF Password Recovery http://www.elcomsoft.com/apdfpr.html
• BarsWF http://3.14.by/en/md5
• My experiments with the GPU, including http://www.golubev.com/rargpu.htm (Although ARCHPR actually also a «my experiment»).

Speed was measured on a single core Intel Core 2 Q6600@2.4Ggts, ATI HD4850 at the nominal frequency 625Mgts, nVidia GTX 260 / w 192SP also at the nominal frequency 1.242Ggts.

Algorithm	Number of blocks	Required cycles	Theoretical speed	Practical speed	ATI HD4850	x	GTX 260	x
RAR 3.x passlen = 4	(4 * 2 + 11) * 4096 + 17 = 77,841 x SHA1	13 622 175	176	160 (ARCHPR) 168 (crark)	3120	19.5	2080	12.4
RAR 3.x passlen = 6	(6 * 2 + 11) * 4096 + 17 = 94,225 x SHA1	16 489 375	145	134 (ARCHPR)	2625	19.5	1695	12.6
WinZip AES	(1000 + 1) * 2 = 2002 x SHA1	350350	6850	6700	135K	20.1	--	--
WPA	(4096 + 1) * 2 * 2 = 16388 x SHA1	2867900	836	820	14К	17	9800	12
MS Office 2007	50005 x SHA1	8750875	274	94	--	--	3300	37.5
PDF9	1 x SHA256	--	--	5.09M	--	--	74.4M	14.6
MD5 single hash	1 x MD5 * (45/64)	51	47M	44.5M	891M	20	563M	12.7

As seen, the ratio of velocity to the GPU to speed on the CPU rather monotonous. ATI HD4850 is about 20 times faster than a single core Q6600 at 2.4GHz frequency, GTX 260 about 12 times. The value of 37.5 for Office 2007 came only because of the non-optimized CPU code (funny that Office 2007 uses SSE2, but so badly that it is better to not use). Speed sorting on the GPU depends only on the frequency of the ALU (as with all the CPU), memory speed and size mean nothing. Knowing the difference in the frequency and number of SP can be fairly accurately assess the performance of different GPU one family.

Compare directly the price the CPU and GPU rather silly - the video card can not work independently, and the CPU too. Systems with two or more processors immediately take us to another level of prices, while there is no economic sense to consider these (server) processors large L2 cache for tasks passwords attack. With GPU is the limiting factor is the number of PCI-E connectors on the motherboard. At the simplest of all, he boards one at most - two. Motherboard with 4 connectors is quite rare. And one must not only 4 slots, but also a place for themselves maps - take something they are usually 2 slot. 4 times 2 slot is found, apparently, only the MSI K9A2 Platinum.There are 6-slot P6T6 WS Revolution, but there can only put a «single» maps.

Seems put together a system with 4 ATI HD4870x2 is impossible in principle (a good description of all the problems here: http://forums.guru3d.com/showthread.php?p=2862271). System with 4 dual nVidia cards, there are ( examplehttp://fastra.ua.ac.be/en/specs.html), but the problems and there enough. Starting with a powerful power supply and case, where everything can be properly placed and finishing with installing drivers.

The overall result is a list of benchmark performance at 4-cores Q6600 at 2.4GHz Frequency:

Title	Frequency, number of kernels	Value	Price CPU or GPU	Total price for the system
E2160	1.8 x 2	0.38x	?	<$ 300
Q6600	2.4 x 4	1x	$ 220	~ $ 600
nVidia 8600 GT	1.188 x 32	0.5x	?	~ $ 650
nVidia 9800 GT	1.242 x 112	1.75x	$ 120	~ $ 720
ATI HD4770	0.75 x 640	5x	$ 100	~ $ 700
ATI HD4850	0.625 x 800	5x	$ 120	~ $ 720
nVidia GTX295	1.242 x 240 x 2	7.5x	500	~ $ 1200
ATI HD4870x2	0.75 x 800 x 2	12x	$ 400	~ $ 1100
4x ATI HD4770	0.75 x 640 x 4	20x	$ 400	~ $ 1200
2x ATI HD4870x2	0.75 x 800 x 2 x 2	24x	$ 800	~ $ 1600
4x nVidia GTX295	1.242 x 240 x 2 x 4	30x	$ 2000	> $ 3000
TACC1441 FPGA	?	2.4x	$ 4000	~ $ 4500

For systems with slow GPU (like 8600 GT) it makes sense to consider performance as a CPU + GPU (that there has not been done). For fast same GPU processor will be loaded with additional calculation and synchronization tasks, so that is additional to it simply makes no sense. In some cases, the quad-core processor may not be sufficient for systems with 8 GPU.

As can be seen, the most interesting solution in terms of price / performance is a bunch of 4-HD4770. Once again, however, requires a motherboard which can accommodate 4 such cards - despite the fact that the HD4770 chip card, the cooling system has an additional slot.

TACC1441 shown for example as an «iron» solutions. This fee is based on FPGA, supported by software from AccessData (http://www.forensic-computers.com/TACC1441.php). As you can see, there was no competition with modern GPU can not make it.

Was left a «little» problem - the almost complete absence of software for the cards ATI. Cards from nVidia are supported in ElcomSoft's Distributed Password Recovery, but with ATI until all very bad.

Conclusion: The increase in productivity from the use of a single GPU approximately 4-5-fold compared with the modern quad system. From ATI noticeably faster at the same price, but there are almost no software. With increasing number of GPU in the system of nonlinearly growing problem - there is a need for a powerful power supply, a special case, efficient cooling system, etc.

[Link]

• http://golubev.com/about_cpu_and_gpu_ru.htm