HD 5870

[BarsMonster]

Get ready for 2500 MHash/sec on a single core

1600 SP
850 Mhz

20 days left

[IvanG]

Get ready for 2500 MHash/sec on a single core

I'm expecting even a bit more

http://www.golubev.com/about_cpu_and_gpu_2_en.htm

[IvanG]

The only report I've got so far is for HD5850. It isn't direct information from HD5850 owner though, so errors possible.

For SHA1 5850 shows 690M (for comparison, it's 290M on 4770), so it's even better than I was expected. However system is pretty unstable, drivers aren't ready at all.

[drewy]

I've got some numbers from a shiny new 5870, they are a little better than predicted.
Will post them a little later.

[BarsMonster]

I guess that's because any of 5 units may do add+any operation. That is extremely useful for cryptography (or actually almost anything).
Finally we've got add+any, not just add+mul

[D3ad0ne]

drewy you f-ing tease!

Oh if you get a chance can you test ighashgpu using a big md5 hashlist? Please.

[drewy]

D:\ighashgpu>IGHASHGPU.EXE /h:cbe1d6d5800ec1e03a5f2a64882a0d41 /t:md5 /c:sd /max
:7
****************************************************************
*** MD4/MD5/SHA1 GPU Password Recovery v0.61 ***
*** For ATI RV 7X0 cards and nVidia 'CUDA' ones (G80+) ***
*** (c) 2009 Ivan Golubev, http://golubev.com ***
*** see "readme.htm" for more details ***
****************************************************************

Found 1 CAL device(s)
Starting brute-force attack, Charset Len = 36, Min passlen = 4, Max passlen = 7
Charset (unicode -> 0) [abcdefghijklmnopqrstuvwxyz0123456789]
Charset in HEX: 61 62 63 64 65 66 67 68 69 6a 6b 6c 6d 6e 6f 70 71 72 73 74 75 7
6 77 78 79 7a 30 31 32 33 34 35 36 37 38 39
Starting from [aaaa]
Hash type: MD5, Hash: cbe1d6d5800ec1e03a5f2a64882a0d41
Device #0: [RV870] 850.00 Mhz 1600 SP
CURPWD: qybcsco DONE: 50.19% ETA: 15s AVRSPD: 2447.3M
Found password: [roger15], HEX: 72 6f 67 65 72 31 35
Processed 41 615 884 288 passwords in 17s.
Thus, 2 461 604 417 password(s) per second in average.

D:\ighashgpu>ighashgpu.exe -h:a72befac3e58eb24d559d9fe0045cfdf090782e2 -t:sha1 -
unicode -max:6 -salt:e16bed51
****************************************************************
*** MD4/MD5/SHA1 GPU Password Recovery v0.61 ***
*** For ATI RV 7X0 cards and nVidia 'CUDA' ones (G80+) ***
*** (c) 2009 Ivan Golubev, http://golubev.com ***
*** see "readme.htm" for more details ***
****************************************************************

Found 1 CAL device(s)
Starting brute-force attack, Charset Len = 26, Min passlen = 4, Max passlen = 6
Charset (unicode) [abcdefghijklmnopqrstuvwxyz]
Charset in HEX: 0061 0062 0063 0064 0065 0066 0067 0068 0069 006a 006b 006c 006d
006e 006f 0070 0071 0072 0073 0074 0075 0076 0077 0078 0079 007a
Starting from [aaaa]
Hash type: SHA1, Hash: a72befac3e58eb24d559d9fe0045cfdf090782e2
Salt: e1 6b ed 51
Device #0: [RV870] 850.00 Mhz 1600 SP
CURPWD: kvhb DONE: 00.18% ETA: 10s AVRSPD: 791.9M
Found password: [carlos], HEX: 0063 0061 0072 006c 006f 0073
Processed 536 870 912 passwords in 1s.
Thus, 903 823 084 password(s) per second in average.


I did say earlier that I had seen faster speeds, and I'm pretty sure I did
I believe the MD5 run above did report 2700, but now it clearly isn't...beta drivers maybe

[drewy]

I'm afraid barswf appears to hang, at least with current ATI beta drivers. I get:

╔═══════════════════════════════════════════════════════════╗
║ BarsWF MD5 bruteforcer v0.9b¨ http://3.14.by/en/md5 ║
║ by Svarychevski Michail http://3.14.by/ru/md5 ║
Ã───────────────────────────────────────────────────────────Â
║ GPU0 : 0.00 MHash/sec CPU0 : 0.00 MHash/sec ║
║ CPU1 : 0.00 MHash/sec ║
║ CPU2 : 0.00 MHash/sec ║
║ ║
║ GPU* : 0.00 MHash/sec CPU* : 0.00 MHash/sec ║
Ã───────────────────────────────────────────────────────────Â
║ Key: M7PM Avg.Total: 0.00 MHash/sec ║
║ Hash:03f75a7c34b81f76bc695014790489bc ║
║ Progress: 51.12 % ETC:.... days .. hours .. min .. sec ║
╚═══════════════════════════════════════════════════════════╝

Which sits there for a few minutes, the box then appears to hang and I get a gpu recovery error.

[drewy]

drewy you f-ing tease!

Oh if you get a chance can you test ighashgpu using a big md5 hashlist? Please.

Of you want to pm me a hashlist, I'll give it a whirl for you.

[BarsMonster]

2447
(c) Just as I predicted

[drewy]

yep, like I say I'm sure I saw 2700 earlier. That's why I posted and said it was a tad faster. But now it's dropped to 2447.
It's now running 2 lcd's and earlier I only had the 1 plugged, I guess it's sorta possible that that would make a small difference.

[D3ad0ne]

drewy this should be a good list of about 9K hashes.
http://hashkiller.com/api/opencrack.php ... e=20091007

I get about 2785M/s with 3 gtx 295s.

[drewy]

what cmd line would you be using for that?

something like: ighashgpu.exe /t:md5 /c:csd /max:6 hash.txt ?

[D3ad0ne]

I would go with...

Code: Select all

ighashgpu.exe -t:md5 -c:csd -min:8 hashlist.txt

min:6 might be to short to get a good average.

[drewy]

I don't think you'll be throwing those 295's away anytime soon
Looking like about 1097 with a 850 core. Increasing the core to 900 only just nudges it over 1100.

[BarsMonster]

Actually we might see more performance as drivers would support new hardware features, because right now it's not.
Officially 5870 is not supported yet AFAIR.

Update: Looks like I am wrong, it is still simple MAD Anyway, it's cool.

[IvanG]

drewy, thanks for the tests.

Results very close to expected though a bit lower. 5870 must be (0.85*1600 / 0.75*640) = 2.83(3) times faster than 4770, so it's should be 2680M for single MD5 & 821M for SHA1. Your SHA1 test runs for less than a second, so result isn't that accurate (it actually differs from 791M to 903M), you can try it with -min:8 and wait for 20-30 sec. Also you can try to add /cpudontcare switch, with such fast GPU it's also a possible reason of problems. One more thing possible, as block size is 2^23 and speed around 2500M it takes only 3.5ms to process one block, so kernel invocation time also can matters here.

Results for hash lists not looking so good, it should be around 1500M not 1100M. Looking at disassembly for RV870 I noticed that there no changes at ALU clauses at all, only memory fetches compiled in different way compared with RV770. As processing hash lists resulting in lots of fetches looks like this is the reason of problem. I guess we need to wait fresh drivers. And, by the way, it's not enough to install new drivers, ighashgpu will also needs recompilation as CAL compiler updates with each Catalyst and GPU kernels prebuild with Catalyst I currently using.

Shortly, I guess ighashgpu will normally works with RV870 only after I'll get it by my own, so it'll be possible to test everything I need. Unfortunately there no signs of 5870s in Russia at all, even at $600 mark.

[IvanG]

5770 will be exactly the same (for hash cracking) as 4890 -- 800SP @ 850Mhz. 5750 as 720SP @ 700Mhz = 4850. But it's RV870, so better architecture anyway .

[BarsMonster]

One more thing possible, as block size is 2^23 and speed around 2500M it takes only 3.5ms to process one block, so kernel invocation time also can matters here.

You might want to implement gpu_time thing for ighashgpu - do a performance benchmark to reach specified kernel execution time

[drewy]

Also you can try to add /cpudontcare switch

made no noticable difference.

[IvanG]

Also you can try to add /cpudontcare switch

made no noticable difference.

I guess I need to increase block size for 5870. Unfortunately algorithm used makes it not as easy thing as it looks.
Actually I'd prefer to use bigger blocks even for 4850 but with 5870 it became mandatory.