Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20121208231803.GA32590@openwall.com>
Date: Sun, 9 Dec 2012 03:18:03 +0400
From: Solar Designer <solar@...nwall.com>
To: john-users@...ts.openwall.com
Subject: Re: GPU based cracking, AMD or NVIDIA

On Sun, Dec 09, 2012 at 01:01:31AM +0200, Milen Rangelov wrote:
> I would say that Kepler-based GPUs are faster than their Fermi counterparts
> (more on "fast" algos and less on "slow" ones).
> 
> There is a reason for that. Generally speaking, the SM/clock speeds are
> better for 6xx. However, Nvidia crippled the architecture in a way. They
> just packed 4x more SPs in a CU, but the register file is just 2x larger.
> This in turn means that on 6xx, running algos that require more GPRs will
> suffer from occupancy problems much more as compared to 5xxx.
> 
> To simplify that: if you are to run "fast" algos like md5 or sha1, 6xx is
> much better (both in MH/s per GPU and MH/watt per GPU). For slower algos it
> might even turn out that 5xx is faster on some occasions.
> 
> You can have a look at my estimations table (based on hardware specs and
> algorithm specifics, not real benchmarks):
> 
> http://www.gat3way.eu/est

As you correctly state, these are merely estimates - not actual results.
The (few) actual benchmark results I've seen so far show GTX 6xx being
significantly slower than it would otherwise have been expected.  Maybe
this was caused by immature drivers/SDK, but I suspect (and this is pure
speculation) that it's also caused by instruction latencies.  That is, I
_guess_ that besides the reduced number of GPRs per SP another problem
is that Kepler's instruction latencies are higher than Fermi's (so we'd
actually need _more_ GPRs to hide the latencies).

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.