[phpBB Debug] PHP Warning: in file [ROOT]/phpbb/session.php on line 580: sizeof(): Parameter must be an array or an object that implements Countable
[phpBB Debug] PHP Warning: in file [ROOT]/phpbb/session.php on line 636: sizeof(): Parameter must be an array or an object that implements Countable
3.14.by forum • 33.1 billion hash/sec
Page 1 of 2

33.1 billion hash/sec

Posted: Thu Dec 09, 2010 3:11 pm
by mrb
Edit: now reaching 33.1 billion password hashes per second... see later comment.

I wrote software and built a machine capable of cracking 28.6 billion MD5 password hashes/sec, handily beating D3ad0ne's challenge and his 17.2 billion hashes/sec :-) My setup is 1.7x faster, consumes half the power (1200W vs. 2000W+), and costs a quarter of his ($2700 vs. $10k+).

This was accomplished with 4 dual-GPU AMD Radeon HD 5970 cards (non-overclocked) on a Linux system, and a new MD5 brute forcing tool, "Whitepixel", that I just released as open source. ighashgpu would NOT have run on my machine because this Windows-only app does not support more than two 5970s due to platform/drivers limitations on this OS. Also, whitepixel is faster than all MD5 brute forcing tools, including ighashgpu.

Picture + details:
http://blog.zorinaq.com/?e=42

It is my turn to ask... who can beat 28.6 Bhash/sec ? :ugeek:

Re: 28.6 billion hash/sec

Posted: Thu Dec 09, 2010 6:03 pm
by D3ad0ne
Nice man, I love a challenge. To bad I can't run this tool on my system.

Also after reading your site you have stats from a 8 GTX580 system. Who has that?

Re: 28.6 billion hash/sec

Posted: Fri Dec 10, 2010 5:28 am
by mrb
Nobody has 8 GTX 580. This is why I wrote "estimated", based on number of ALUs and ALU clock.

Has anyone on the forum tried my method to down-plug a PCIe x16 graphics card in an x1 slot?

Re: 28.6 billion hash/sec

Posted: Fri Dec 10, 2010 5:13 pm
by D3ad0ne
I been able to use a x1 slot with an x16 card before. I haven't tried your method though.

There has been some talk about your application. If you get a chance you should join us in the #hashcat channel on irc.rizon.net

Re: 28.6 billion hash/sec

Posted: Sat Dec 11, 2010 9:36 pm
by P3ndldqny2
Are we suggesting, if someone had 8x Radeon 6790 cards we could get roughly 80billion/second on this Linux software?

I'm presuming 8x 6790s would be better than 8x GTX 580s ?

Re: 28.6 billion hash/sec

Posted: Sun Dec 12, 2010 8:48 pm
by Sc00bz
I'm going to guess you mean Radeon 6970 not Radeon 6790.

A Radeon 6970 58.3% the speed of a Radeon 5970 (assuming the stats of 880 MHz and 1536 cores are correct). So 8 Radeon 6970 would give you 33.3 billion/second (assuming they can do the same amount of work per clock-core).

Re: 28.6 billion hash/sec

Posted: Mon Dec 13, 2010 1:10 am
by mrb
Scoobz is correct (as usual).

But expect a nice surprise with whitepixel v2 :wink:

Re: 28.6 billion hash/sec

Posted: Mon Dec 13, 2010 11:10 pm
by P3ndldqny2
Sc00bz wrote:I'm going to guess you mean Radeon 6970 not Radeon 6790.

A Radeon 6970 58.3% the speed of a Radeon 5970 (assuming the stats of 880 MHz and 1536 cores are correct). So 8 Radeon 6970 would give you 33.3 billion/second (assuming they can do the same amount of work per clock-core).
I assume the 6970 is faster than the gtx 580?

Re: 28.6 billion hash/sec

Posted: Tue Dec 14, 2010 5:21 am
by Sc00bz
P3ndldqny2 wrote:
Sc00bz wrote:I'm going to guess you mean Radeon 6970 not Radeon 6790.

A Radeon 6970 58.3% the speed of a Radeon 5970 (assuming the stats of 880 MHz and 1536 cores are correct). So 8 Radeon 6970 would give you 33.3 billion/second (assuming they can do the same amount of work per clock-core).
I assume the 6970 is faster than the gtx 580?
Oops I forgot to answer that, yes.

Re: 28.6 billion hash/sec

Posted: Tue Dec 14, 2010 1:24 pm
by mrb
I released whitepixel v2. The nice surprise is... support for the BFI_INT instruction which allows implementing the MD5 subfunctions F() and G() in 1 clock cycle. My 4x5970 setup now reaches 33.1 billion passwords per second.

http://blog.zorinaq.com/?e=43

Image

Re: 28.6 billion hash/sec

Posted: Wed Dec 15, 2010 11:36 pm
by P3ndldqny2
mrb wrote:I released whitepixel v2. The nice surprise is... support for the BFI_INT instruction which allows implementing the MD5 subfunctions F() and G() in 1 clock cycle. My 4x5970 setup now reaches 33.1 billion passwords per second.

http://blog.zorinaq.com/?e=43

Image
What do you reckon 8x 6970s would reach, with say a dual xeon setup (24 hyperthreads?)? (I presume 8x is the maximum of graphics cards people can connect, or is it 7?)

Re: 33.1 billion hash/sec

Posted: Thu Dec 16, 2010 9:51 am
by mrb
8xHD 6970 should do about 38.6 Bhash/sec under whitepixel v2.
Possibly up to ~41 Bhash/sec with some more code optimization ideas I may have...

Linux drivers support at least 8 GPUs. Windows drivers are limited to 4 it seems.

CPU speed is irrelevant for whitepixel. It does not use the CPU at all. CPU time stays at a constant 0% on my 4x5970 machine churning at 33.1 Bhash/sec...

Re: 33.1 billion hash/sec

Posted: Fri Dec 17, 2010 12:34 am
by P3ndldqny2
mrb wrote:8xHD 6970 should do about 38.6 Bhash/sec under whitepixel v2.
Possibly up to ~41 Bhash/sec with some more code optimization ideas I may have...

Linux drivers support at least 8 GPUs. Windows drivers are limited to 4 it seems.

CPU speed is irrelevant for whitepixel. It does not use the CPU at all. CPU time stays at a constant 0% on my 4x5970 machine churning at 33.1 Bhash/sec...
I hate to sound thick here, but if 4x5970 = 33.1 billion, why doesn't 8x6970 = 70billion(ish)?

Re: 33.1 billion hash/sec

Posted: Fri Dec 17, 2010 5:10 am
by Sc00bz
HD5970 has two GPUs in it.

4 HD5970 * 725 MHz * 1600 cores * 2 GPUs = 9,280,000
vs
8 HD6970 * 880 MHz * 1536 cores = 10,813,440

8 HD6970 is 1.1652 times faster than 4 HD5970.

Re: 33.1 billion hash/sec

Posted: Sat Dec 18, 2010 5:44 pm
by P3ndldqny2
Sc00bz wrote:HD5970 has two GPUs in it.

4 HD5970 * 725 MHz * 1600 cores * 2 GPUs = 9,280,000
vs
8 HD6970 * 880 MHz * 1536 cores = 10,813,440

8 HD6970 is 1.1652 times faster than 4 HD5970.
Ah ok thanks for this. I (wrongly) made the assumption that each new ATI GPU would be better than the previous, but obviously not the case with the 5970.

So, 8x 5970s could produce about 66 billion hashes/second?

Re: 33.1 billion hash/sec

Posted: Sat Dec 18, 2010 5:52 pm
by BarsMonster
Just a few thoughts: The lower GPU clock/voltage, the moar GFLOP/W you get.
That's why dual-chip GPU rock so hard: due to lower clock AND voltage each operation costs less energy.

My tests shows that on half the clock you spend 5 times less power per operation.

Re: 33.1 billion hash/sec

Posted: Sat Dec 18, 2010 10:13 pm
by mrb
I highly favor performance/Watt as opposed to absolute performance, so, yes, I am planning to do many undervolting and underclocking tests... However I am pretty sure that undervolting is more effective than underclocking. My preliminary tests so far show that decreasing the core frequency of one of my 5970s from 725MHz to 550MHz with aticonfig decreases performance by 24% while only decreasing power consumption by 18%: 22.1A@12V down to 18.1A@12V (overall current draw on PCIe slot plus 6-pin and 8-pin connectors), effectively decreasing perf/W.

P3ndldqny2: correct. One guy who contacted me said he was going to try 8x5970. We will see if the drivers support that...

Re: 33.1 billion hash/sec

Posted: Mon Dec 20, 2010 10:58 am
by mrb
Mysql323 is kind of low on my priority list because it is rarely requested.

I have also heard (IIRC) of performance levels of trillions of passwords per second with CPU-based mysql323 brute forcers, so this symbolic barrier has already been broken :-)

Re: 33.1 billion hash/sec

Posted: Mon Dec 20, 2010 5:52 pm
by BarsMonster
mrb wrote:However I am pretty sure that undervolting is more effective than underclocking. My preliminary tests so far show that decreasing the core frequency of one of my 5970s from 725MHz to 550MHz with aticonfig decreases performance by 24% while only decreasing power consumption by 18%: 22.1A@12V down to 18.1A@12V (overall current draw on PCIe slot plus 6-pin and 8-pin connectors), effectively decreasing perf/W.
Surely you need to combine :-D
My tests were showing like 80-90% less power at 400Mhz and 0.85v (to go below 1v you need to lower ram frequency too)

Re: 33.1 billion hash/sec

Posted: Tue Dec 21, 2010 7:35 am
by Sc00bz
The CPU code that does over a trillion passwords per second for MySQL323 has divergent code so it will be hard to get nice performance with GPUs or SSE2. Also it only works for a single hash at a time so multiple hashes will slow it down until it gets to a point where it makes more sense to just generate hashes and compare them against a list of hashes you are cracking.

Anyway it might get just over 1 quadrillion (my logic: 1 trillion * 2 SIMD/SISD * 28,600/50 GPUs/Core2Solo) but I doubt anyone can get it that optimized.

Re: 33.1 billion hash/sec

Posted: Wed Dec 22, 2010 12:33 am
by P3ndldqny2
mrb wrote:Mysql323 is kind of low on my priority list because it is rarely requested.

I have also heard (IIRC) of performance levels of trillions of passwords per second with CPU-based mysql323 brute forcers, so this symbolic barrier has already been broken :-)
What is this method you speak of? I tried google but didnt get much

Re: 33.1 billion hash/sec

Posted: Wed Dec 22, 2010 3:20 am
by mrb
P3ndldqny2 wrote:
mrb wrote:Mysql323 is kind of low on my priority list because it is rarely requested.

I have also heard (IIRC) of performance levels of trillions of passwords per second with CPU-based mysql323 brute forcers, so this symbolic barrier has already been broken :-)
What is this method you speak of? I tried google but didnt get much
Scoobz describes the algorithm to compute the last 2 chars of a mysql323 pw hash in more details here:
http://freerainbowtables.com/phpBB3/vie ... a&start=60
So technically speaking the last 2 chars are not brute forced, but from an external viewpoint, it is as if the brute forcer was able to instantly try all possible combinations of the chars, hence the equivalent of "trillions/sec".

Re: 33.1 billion hash/sec

Posted: Fri Dec 24, 2010 12:45 am
by knapish
ati 6990 = 1920*2*~900=13824000
With a price of ~600$ wont actauly be that funny anymore... :wink:

Re: 33.1 billion hash/sec

Posted: Fri Dec 24, 2010 5:11 pm
by Rolf
Sc00bz wrote:Also it only works for a single hash at a time so multiple hashes will slow it down until it gets to a point where it makes more sense to just generate hashes and compare them against a list of hashes you are cracking.
I have done extensive tests on SchwarzwaldMySQL(the only CUDA-enabled software to bruteforce MySQL hashes with that method), and here are some of conclusions:
If you increase the amount of hashes(which must be > 0 ) twice, you will get a ~3.3x speed drop.
Bruteforcing a hash at a time is FASTER than bruteforcing all the hashes, with the exception that you find a lot of hashes at the start, so the process will speed itself up.
It is rendered ineffective at 256 hashes.

It wont perform on AMD cards as good as on NV cards because of vector datatype.

Re: 33.1 billion hash/sec

Posted: Thu Feb 24, 2011 1:27 am
by Forsaken
mrb wrote:... However I am pretty sure that undervolting is more effective than underclocking.

Power consumption goes up linearly with frequency and quadratically with voltage. For a CMOS device P = CV^2f.


mrb do you have a forum for the whitepixel project? Ive hit some compile errors and could use some reference.