Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20131203222651.GA24951@openwall.com>
Date: Wed, 4 Dec 2013 02:26:51 +0400
From: Solar Designer <solar@...nwall.com>
To: john-users@...ts.openwall.com
Subject: Re: Energy-efficient bcrypt cracking (Passwords^13 slides)

On Tue, Dec 03, 2013 at 10:31:12PM +0100, Lukas Odzioba wrote:
> I've got two questions to the slides. Are you sure your power estimations
> for gpus are correct?

As estimates, yes.  They're not exact, but they're good enough.

> What metodology did you use?

It was tricky.  For these measurements, we used GPUs in two machines:
bull (HD 7970 and GTX 570) and super (HD 7990 and GTX TITAN).  For bull,
I was measuring its total power consumption from the 220V AC socket, so
we had to rely on the idle-to-load deltas and on the vendors' published
power consumption figures for these GPUs at idle.  For super, I obtained
deltas for both AC and 12V DC via IPMI, which in turn got them from
super's two PSUs via PMBus.  bull's PSU is officially 85% efficient.
super's PSUs officially have over 94% efficiency, and there was in fact
not much difference between AC and 12V DC deltas for GPU load.

For Xeon Phi, also in super, we had more data: the Xeon Phi card has its
own SMC, so it reports its own full power usage (and indeed we obtained
the delta too).  This is in addition to what we got from the PSUs.  So
for every benchmark we ran, we could obtain three separate power
measurements (from two SMCs for DC, and for AC).  These pretty much
matched (all three deltas were in the range 35 W to 40 W).

> Values for gpus looks a bit too small for me.

Actually, before we started measuring, I expected them to be even
smaller, because bcrypt uses at most 1/4th of HD 7970's computing power.
My guess was "probably way below 100 W"; in reality, it's around 100 W
and sometimes even more than that.

> Especially when "device power estimate" is less than "iddle to load delta".

As this slide says:

http://www.openwall.com/presentations/Passwords13-Energy-Efficient-Cracking/slide-34.html

"System power consumption includes PSU overhead (typically 5% to 20%),
hence deltas may exceed CPUs' TDP"

The same holds true for GPUs (next slide), and it applies to deltas vs.
device power estimates.

> It looks like it does not include
> some iddle load which may be revelant, or this value is defined in
> some other way.

Here's the calculation for HD 7970:

15 W idle + 95 W delta * 0.85 PSU efficiency = ~96 W

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.