hashcat Forum
Wordlist with rule generation (-g), performance decrease in 3.10 compared to 2.01 - Printable Version

+- hashcat Forum (https://hashcat.net/forum)
+-- Forum: Support (https://hashcat.net/forum/forum-3.html)
+--- Forum: hashcat (https://hashcat.net/forum/forum-45.html)
+--- Thread: Wordlist with rule generation (-g), performance decrease in 3.10 compared to 2.01 (/thread-5854.html)



Wordlist with rule generation (-g), performance decrease in 3.10 compared to 2.01 - Arlaine - 09-10-2016

I understand there's a bit of a performance decrease to be expected when using rules+wordlists on some hash types in hashcat v3.00+ ( as per atom's comments here ) but I wanted to verify if the performance hit should be as large as what I'm seeing or if there is a misconfiguration on my part.


The commands being run are as follows [truncated]:
Code:
oclHashcat-2.01>oclHashcat64.exe --gpu-temp-disable -m 1000 hashes.txt rockyou.txt -g 1000000
hashcat-3.10>hashcat64.exe --gpu-temp-disable -m 1000 hashes.txt rockyou.txt -g 1000000

These have been tested with "-w 3" and some light fiddling but the results always end up quite far apart in performance. [screenshot link below, also attached]
Left: oclHashcat-2.01
Right: hashcat-3.10
http://i.imgur.com/IDgHdlw.png

.png   hashcat_2v3_wordlistPlusRules.PNG (Size: 28.85 KB / Downloads: 9)

Everything else is in line with other comments on the forums where 3.00+ is faster in both bruteforce and wordlist-only modes.
If there are some commands I can run to provide additional information or test particular parameters please let me know and I'll gladly throw them up.


RE: Wordlist with rule generation (-g), performance decrease in 3.10 compared to 2.01 - kartan - 09-10-2016

You can't expect -g to give you the same performance every time simply because it randomly generates new rules at every start.

Different rules -> different performance

To benchmark rules take a fixed ruleset.

Tho, that's quite a difference between the version.


RE: Wordlist with rule generation (-g), performance decrease in 3.10 compared to 2.01 - kartan - 09-10-2016

I've just tested it on my 980ti and for me v2.01 is only half the speed of v3.10

Your observation seems to be very specific to your system.

What kind of GPUs do you have there and which driver version do you have running?


RE: Wordlist with rule generation (-g), performance decrease in 3.10 compared to 2.01 - atom - 09-10-2016

We said that like 1000 times here and on the release notes. Pre-Maxwell GPU will suffer a performance drop, but new GPU will be faster with v3 than with v2. For all Pre-GCN GPU it's the same (5xxx and 6xxx). It's the future that matters, we optimize for that.


RE: Wordlist with rule generation (-g), performance decrease in 3.10 compared to 2.01 - Arlaine - 09-10-2016

Thank you both for the answers and taking the time; and sorry on the late response, I wanted to try to test something on another setup.

(09-10-2016, 01:07 AM)kartan Wrote: You can't expect -g to give you the same performance every time simply because it randomly generates new rules at every start.
Different rules -> different performance
To benchmark rules take a fixed ruleset.
Tho, that's quite a difference between the version.

Agreed, normally I'd take this route but the variance observed in a given hashcat version on 1million generated rules run-to-run is within a few seconds difference so I'm not too worried about it for the current scenario.


(09-10-2016, 01:23 AM)kartan Wrote: I've just tested it on my 980ti and for me v2.01 is only half the speed of v3.10

Your observation seems to be very specific to your system.

What kind of GPUs do you have there and which driver version do you have running?

My bad, the setup is:
3x AMD R290X
Windows server 2012r2
Driver version 16.150.2211.0 / Catalyst 15.8


(09-10-2016, 10:15 AM)atom Wrote: We said that like 1000 times here and on the release notes. Pre-Maxwell GPU will suffer a performance drop, but new GPU will be faster with v3 than with v2. For all Pre-GCN GPU it's the same (5xxx and 6xxx). It's the future that matters, we optimize for that.

Understood, and that's easily reflected in any test of the two versions under most circumstances, which is why it felt strange to see the variance I'm seeing, again this may just be an issue with my setup.

I tried on an RX480 on hashcat v3.10 as well, running the same test as the initial post, I got roughly:
750 MH/s
which is still quite far off from v2.01 on the initial test.


I have no issues with the performance on this one attack mode being different, like you've said its about supporting for the future. I use the old version only when I need to run this specific attack type anyways as v3.00+ is better in every other way.
It just seemed like a fairly large difference in this instance and was curious to hear thoughts on it, as with most things that seem odd there's usually operator error of some sort