[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20160709201744.GA31206@openwall.com>
Date: Sat, 9 Jul 2016 23:17:44 +0300
From: Solar Designer <solar@...nwall.com>
To: john-users@...ts.openwall.com
Subject: Re: Nvidia GTX 1070 md5crypt-opencl performance
On Fri, Jul 08, 2016 at 02:40:36PM -0400, Andy Lawrence wrote:
> First time JtR user, so I could have this messed up, but thought maybe some
> would be interested in the new Nvidia performance:
Yes, thanks!
> Loaded 1 password hash (md5crypt-opencl, crypt(3) $1$ [MD5 OpenCL])
> Node numbers 1-2 of 2 (fork)
> Device 0: GeForce GTX 1070
> Device 1: GeForce GTX 1070
> Press 'q' or Ctrl-C to abort, almost any other key for status
> ~snip~
> 1 0g 0:00:53:25 3/3 0g/s 4464Kp/s 4464Kc/s 4464KC/s GPU:74C
> dyomuro2*..dyokayn3s
> 2 0g 0:00:53:25 3/3 0g/s 4203Kp/s 4203Kc/s 4203KC/s GPU:68C
> 09801angi..09807buli
For md5crypt, performance varies significantly by password length being
tested. Here we see JtR testing length 9, but the speeds are averages
over the 53 minutes during which other lengths were being tested as
well. So while this sort of benchmark may reflect real-world runs well,
it is not convenient for testing one GPU vs. another, etc., where it'd
be more convenient to run a shorter benchmark for a specific length.
md5crypt performance may also vary by salt length. It's usually 8, but
it can be anywhere from 0 to 8.
Here's Titan X at stock clocks attacking one md5crypt hash with salt
length 8 and with "-inc -min-len=8 -max-len=8", so testing 8 character
long candidate passwords only:
0g 0:00:01:00 0g/s 4792Kp/s 4792Kc/s 4792KC/s GPU:72C letry21b..leileptu
For comparison, here's length 7 ("-inc -min-len=7 -max-len=7"):
0g 0:00:01:00 0.00% (ETA: 2017-01-02 23:03) 0g/s 4527Kp/s 4527Kc/s 4527KC/s GPU:72C k18rmum..k1t06sd
and length 9:
0g 0:00:01:00 0g/s 4451Kp/s 4451Kc/s 4451KC/s GPU:82C l00672032..l00656701
Driver 361.28. Your speeds suggest that the Pascal cards certainly are
faster (considering that yours are 1070's and not 1080's), as expected.
There's also much room for optimization through per-length md5crypt
OpenCL kernels, like latest Hash Suite has. We should consider that.
> Very open to suggestions if my method above is poor.
It's reasonable, but it's hard to use it for exact comparisons of GPUs.
Alexander
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
