[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CABh=JRHa+N4xXT2G+nm+OkMibubsMRcMTWr9bMQwnZa6jd_FFQ@mail.gmail.com>
Date: Sun, 9 Dec 2012 01:49:12 +0200
From: Milen Rangelov <gat3way@...il.com>
To: john-users@...ts.openwall.com
Subject: Re: GPU based cracking, AMD or NVIDIA
I have a Kepler card myself (GT640) that I bought just to experiment. I
also have a sm21 one (GT430). I can tell for sure the MD5 and SHA1 numbers
in the table are pretty close to what I really get. Yet, the table is
misleading. With slow algos that have a lot of GPR utilization/shared
memory utilization, Kepler behaves just bad. Notable example is
sha512-crypt which (in my bad implementation) uses a lot of shared memory
and registers. Currently GT430 is about as fast as GT640 while the latter
has 4x more SPs (though clocked at 900MHz as compared to GT430 which runs
at 1400MHz).
On Sun, Dec 9, 2012 at 1:34 AM, magnum <john.magnum@...hmail.com> wrote:
> On 9 Dec, 2012, at 0:19 , Solar Designer <solar@...nwall.com> wrote:
> > On Sat, Dec 08, 2012 at 11:40:38PM +0100, magnum wrote:
> >> On 8 Dec, 2012, at 19:57 , Solar Designer <solar@...nwall.com> wrote:
> >>> Are you aware of any password cracking related task where GTX 680 is
> any
> >>> faster than GTX 580? I've only seen (non-JtR) benchmarks showing it
> >>> being slower - often several times slower, even. (I don't have one to
> >>> test myself, though.)
> >>
> >> With late OpenCL drivers the gap has decreased a lot, or the 680 has
> even become the winner at times - with a really tiny margin. I have never
> seen the 680 significantly outperform a 580 in any application. Like you, I
> have seen the opposite but I think that was mostly due to premature drivers
> that did not handle Kepler well.
> >
> > This is interesting. Can you post any specific benchmark results?
>
> Will do. The only concrete example I have from memory is NTLMv2, which is
> not a good comparison because it is severaly limited by transfer speed.
> Anyhow, I saw the Kepler win with what looked "significant" to my eye (but
> that might be just above 19M vs just below 18M, which is not a great deal)
> and the two machines were identical except for the GPU cards. This fits
> with what Milen says.
>
> But I am also pretty sure I have seen "slow" formats perform just a tiny
> bit better with Kepler at slow formats, with decent drivers. I'll collect
> some statistics next time I get a chance.
>
> Regardless of what single-digit percentage boosts I *may* report later,
> this is silly. A GTX680 released 2012 should be significantly faster than a
> GTX580 from 2010. Someone please explain Moore law to nvidia :-)
>
> magnum
>
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.
