[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20120329014904.GF23961@openwall.com>
Date: Thu, 29 Mar 2012 05:49:04 +0400
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: Working on DES format on CUDA
On Wed, Mar 28, 2012 at 09:02:44AM +0200, Simon Marechal wrote:
> Just a note : I know hashcat author, Atom, tried a bitsliced DES
> version, but reported that the register usage was too high to make it
> competitive with the "vanilla" version. I thought this was surprising
> but he is probably the most experienced GPU-cracker writer around. This
> means it is probably non trivial.
It is definitely not trivial, but another person who implemented DES on
GPU - chapuni on #openwall - said that bitslice was faster despite of
the register pressure (and reported speeds for 5970 that were much
higher than what hashcat now achieves on 7970 according to its published
benchmarks). (Of course, 5970 is simply faster than 7970, but the
reported speed differences were much greater than that.) That's for a
hard-coded salt value, though, so we'd need to have 4096 precompiled
kernels or to binary-patch them at runtime. BTW, this same
optimization would achieve about a 7% speedup on CPU compared to what we
currently have (I tried it briefly).
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.
