More speed estimations

I’m just realized that I’m performing the same calculations all over again — how bad or good this GPU comparing to that one. It’ll be much easier to compile everything together just once and reuse. So, here we go:

(5830 & Fermi estimations updated at March, 19).

Some notes:

1. Everything is just estimations. Although for GTS250, GTX260, HD4770, HD4850, HD5770, HD5870 & HD5970 real single MD5 & SHA1 hashing speed is very close to these estimations (like ±1%).

2. Estimations for MS Office 07, WinZip/AES & WPA done for perfect conditions, i.e. there will be some additional calculations in real life, so performance will drop by some amount. However for Office 2007 files (which just using 50 000 of SHA1 iterations) real life results also very close to estimations (±5%).

3. Absolutely pointless to apply above estimations to anything else than MD5 & SHA1 hashing (like 3D games or SETI@Home or …).

4. ATI Stream SDK quality is still very questionable, it’s much harder to program ATI GPUs to get peak performance. And it’s almost impossible to do with OpenCL — low level CAL/IL required. Don’t expect a lot of software for ATI Stream in near future.

5. Clock rates for Fermi are still unknown, I’ve just used 1500Mhz which looks reasonable (for me). As Fermi specifications finally published (March, 19 2010) I’ve updated clocks to 1401 & 1215 Mhz and recalculated all other values.

6. I’ve used for price estimations.

7. With all above notes in mind — HD 5770/512MB currently is most cost efficient GPU for MD5/SHA1 hashing.

This entry was posted in GPU programming, Hash cracking and tagged , , . Bookmark the permalink.

38 Responses to More speed estimations

  1. dan says:

    Thanks Ivan, this chart will help me decide which GPU I’m going to buy. :)

  2. Pingback: GPU comparison by Ivan Golubev « iXpLiZiT

  3. iXpLiZiT says:

    Thanks for this nice chart. It seems that the nvidia cards are expensive and slow comparded to the ati cards.

    ATI 4870:
    Single md5: 1050M Price: $170

    nvidia geforcegtx 260:
    single md5: 580M Price: $180

    • Ivan Golubev says:

      Looking from end-user’s side — yes. Looking from programmer’s side — it takes a week to design, code & debug ATI kernel while the same thing takes just 30 minutes with nVidia GPUs.
      Still wonder why so few ATI Stream software around? ;)

  4. FloW3184 says:

    ivan – you seem to know a lot about the lowlvl-programming of ati hardware… would it be possible for you to look into the source of an openCL using emulator? … an open source gamecube-/wii-emulator. because it seems to have issues with ati cards using openCL for decrypting textures on the fly while playing. (without opencl dolphin uses the CPU to decrypt the textures)

  5. FloW3184 says:

    something else…:
    i just bought a 5770 and now i cant use ighashgpu 0.70bt for some hashes…

    for some it is no problem… my nickname in md5 for example… or roger15
    others are not working at all anymore.
    E:\Dinge\Systemtools\ighashgpu>D:\Dinge\Systemtools\ighashgpu\ighashgpu.exe /h:9d218d5b27e1a91399d8488a5569b506cd6b7e01 /t:mysql5 /c:sd /max:10 /sf:6eb3ahzih
    Das Gerät ist nicht bereit. (mashine not ready)

    Drücken Sie eine beliebige Taste . . .

    or md5:
    E:\Dinge\Systemtools\ighashgpu>D:\Dinge\Systemtools\ighashgpu\ighashgpu.exe /h:5295e1c13302f6a462977a1522330788 /t:md5 /c:d /max:8
    Das Gerät ist nicht bereit.

    Drücken Sie eine beliebige Taste . . .

    what could be the problem?

    i am using catalyst 10.2 on win7 x64

  6. Ivan Golubev says:

    I doubt this “Device not ready” error produced by ighashgpu itself (as well as hash value should be irrelevant). It more looks like a problem with device you’re starting ighashgpu from. May be it’s damaged SSD/HDD, may be something else with OS.
    Anyway, I don’t have any problems with Catalyst 10.2 & Vista64.

  7. FloW3184 says:

    it only happens with some hashes… but always the same hashes… if a hash worked once… it works always… if a hash does not work – it never does. (but no problem with 0.62 – working with all hashes i tested)

    • Ivan Golubev says:

      FloW3184, that’s weird. There weren’t any major code changes since 0.62 (except GPU kernels code) and so I have no idea what’s the reason of this “Device not ready issue”.

  8. forester says:

    hi, ATi Radeon HD 4670 with 1GB DDR2 is supported?

  9. forester says:

    i thinking about radeon mobility HD4670

    • Ivan Golubev says:

      Should works. Though usually it’s a problem to install latest drivers for notebooks, some tweaking needed.

  10. forester says:

    what chance I have when I will use ATI Radeon HD5650 ? this card arent on the list.

    • Ivan Golubev says:

      5650 supported by ighashgpu and not supported (as well as all all other non 58XX GPUs) by igrargpu.

      I’m not planning any updates for ighashgpu in near future — simply have no time for this.

      it’s really hard to say what’s the reason of this “Device not ready” issue. Obviously best choice is to reinstall whole system :). As well as most time consuming one. But you can start from reinstalling ATI drivers — uninstall them completely (though sometimes it isn’t that easy) and get 10.2. Although still there are enough problems with 5XXX series — recently I got report (and checked it by my own) that running ighashgpu and playing video with Windows Media Player (and probably any other players which using video output in “advanced” way) results in whole system lock-up == hello, my lovely reset button…

      with my 5770 I only have sometimes gray lines at random positions at screen. But looks like it’s common issue with 57XX GPUs, no other artifacts. Yet :D.

  11. FloW3184 says:

    i made a misstake: in 0.62 i have the same error message about “device not ready” and the same hashes that work and/or not work.

    could it have something to do with my previous radeon hd 4770?
    some driver faults?

    or maybe the ati stream sdk i have installed?

  12. FloW3184 says:

    for better estimations:
    gtx470’s specifications are now known.
    shader clock: 1255 MHz
    ram clock: 1600 MHz (GDDR5 800Mhz)
    not known is how many shader cores are usable on gtx470 though… (at least out of this text…) i think 448.

    for gtx480 it isnt know too… maybe they will cap it to 480 cores.
    final specifications will be sent out to the board producers on next monday.

  13. ksj says:

    yeah, and 5830 is 1120/750

    PS My asus 5770 now have artifacts all the time – probably faulty card from the start, lasted for 3 months only.

  14. FloW3184 says:

    ok, i uninstalled everything…
    driver (ccc & everything else from ATI), Stream SDK, DX SDK, Windows SDK, visual studio express, virus scanner… no chance… same issue… with a nearly empty computer.

    if you are interested contact me via email and i can let you connect to my computer via teamviewer or realvnc to look after that issue. (but i think you won’t, because you stopped developement)

    • Ivan Golubev says:

      FloW3184, I’ve stopped ighashgpu development but I’m still programming GPUs. Anyway, I simply have no idea what’s happens so remote connection kinda pointless for me. You can try to run other ATI Stream applications and check out are there any similar issues this them. From programming point of view different hashes processed in absolutely same way (of course if you’re starting program with same brute-force settings), so there should be no difference at all. Weird.

  15. KT says:

    Do you have pay-per-use service where you help to unlock RAR files? I need to unlock RAR archive and looking for a professional service that can do it for a fee. Please let me know. Thanks

    • Ivan Golubev says:

      KT, I don’t have such service. Generally, if password was well chosen (not simple word from dictionary and longer than 7 symbols) there no chances at all to crack (unlock, recover or any else “good” verb here) it — it’ll require millions of years.

  16. iXpLiZiT says:

    Oh I see. The programming of ati takes a long time. But at the end the ati card will have double, triple, quad or more speed as a nvidia card for the same money or even for a cheaper price.

    • Ivan Golubev says:

      Well, it isn’t that straightforward.
      If algorithm implementable on GPU and it’s (mainly) ALU bound and you have strong desire to make it run as fast as possible and you have enough free time to learn (very poorly documented) ATI hardware and you have enough will to fight through all ATI SDK bugs… then, yeah — you’ll get better speed on ATI GPUs with cheaper price of them.
      But if you’re simply taking some OpenCL code written for nVidia GPU and compiling it for ATI GPU in most cases you’ll end with absolutely terrible performance. To get good speed at ATI GPUs you’ll need to tune your code specifically for ATI hardware while nVidia hardware isn’t that affected executing “badly written” code.

  17. ksj says:

    Ivan, if you didn’t read it yet:
    there’s radeon 5XXX now.

    • Ivan Golubev says:

      Yeah, I’ve seen it. Looks like elcom hasn’t used bitalign as speed for 5970 about 35% lower than my predictions.
      Probably I should finally release igwpapmkgen :).

  18. FloW3184 says:

    omg… you are really working on a wpa password-bruter?
    seems like you are trying to make elcomsoft jobless. xD

    what will be needed to use it?
    will it have a built in access to the program whose name i can’t remember right now… (for reading the “raw data” from wlans)

  19. FloW3184 says:

    tcpdump was its name ;)

  20. Jonny says:

    Hello, ighashcpu don’t work with an ATI READON HD 5650! Driver are still up to date, can please somebody help? Using Windows 7 Home.

    from ighashcpu
    Found 1 CAL device(s)
    Cannot find any compatible GPU device.

    Dear Ivan maybe you can FIX or ADD this Readon Model to ighashcpu?

    Regards, Jonny

    • Ivan Golubev says:

      Post the exact command line you’re using. Also add /debuglog to it, it should shows something like “CAL device 0, target = 9″.

      • tnkflx says:

        I have the same problem…

        Exact command line: “ighashgpu.exe /h:cbe1d6d5800ec1e03a5f2a64882a0d /t:md5 /c:sd /max:7 /cpudontcare /debuglog”


        *** MD4/MD5/SHA1 GPU Password Recovery v0.62 ***
        *** For ATI RV 7X0 cards and nVidia ‘CUDA’ ones (G80+) ***
        *** (c) 2009 Ivan Golubev, ***
        *** see “readme.htm” for more details ***
        *** Any commercial use of this program is strictly forbidden ***

        Found 8 CAL device(s)
        CAL device 0, target = 15
        CAL device 1, target = 15
        CAL device 2, target = 15
        CAL device 3, target = 15
        CAL device 4, target = 15
        CAL device 5, target = 15
        CAL device 6, target = 15
        CAL device 7, target = 15
        Cannot find any compatible GPU device.

        Any idea?

  21. Per says:

    What is best between

    IBM THinkpad t410
    Intel Core i7-620M 2,66GHz
    256MB nVIDIA Quadro NVS 3100M

    Alienware m11x
    Intel® Core™ i7 640UM (4M Cache, 1.20 GHz
    GB GDDR3 NVIDIA® GeForce® GT 335M

    • Ivan Golubev says:

      Both systems not good for serious computations. However NVS 3100M is way worse than GT 335M comparing the peak performance.

  22. Per says:

    how many cracks per second could I get with each laptop?

    with the ibm or the alienware.

    And what abotu the intel core i7 640um 1.2 ghz vs core 7 620m 2,66ghz

    So I guess if each system use cpu+gpu how many cracks total per system?

    Like ibm 8000 cracks per second
    Alienware 9000 cracks per second?

    • Ivan Golubev says:

      ~57M single MD5 for NVS 3100 & ~190M for GT335M. Processors rating is out of scope.

  23. Per says:

    Oh so the NVS 3100 is REALLY bad I can see.

    Do you know for wpa cracking how many cracks per second for each?

    I see a cheaper Dell Vostro with i7 also.
    It has a graphics card: : 512 MB NVIDIA® Geforce™ 310M
    is this card worse or better?

    • Ivan Golubev says:

      Actually you can make all these estimations by your own — open this link:
      310M having 16 cores at 1.53Ghz speed, thus peak performance 16*1.53=24.48. Take any estimation from above tables for GeForce 2XX/3XX family and scale it with this number.
      Generally — the whole NVIDIA 3XX line is very-very weak. Especially comparing with Mobile Radeons, for example, 5730M exactly matches 5570 and that means it at least 10x times faster than NVIDIA GF 310M.

      My personal favorites among notebooks — ASUS N61Ja/N71Ja.

  24. akipfer says:

    can i use my XFX HD6850 for password cracking?
    also, why rar cracking is so slow compared to MD5? i see that MD5 can do millions of, but rar can’t do that, why is that?
    would be great if u make a vid to teach us how to use the gpu for all of this!
    btw good software dude

    • Ivan Golubev says:

      I guess it’s better to use oclHashcat with your HD6850 because ighashgpu become way too outdated for most modern GPUs.

      As for RAR archives — it uses algorithm called “key stretching”, i. e. hashing input password several thousand times, thus millions became only thousands for RAR. Not mentioning other problems with it like whole file unpacking which requires complete UnRAR functionality within password cracker program.

Comments are closed.