Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sun, 7 Oct 2012 23:18:23 +0400
From: Solar Designer <>
Subject: Re: Password hashing at scale (for Internet companies with millions of users) - YaC 2012 slides

On Sun, Oct 07, 2012 at 04:56:52PM +0200, magnum wrote:
> On 5 Oct, 2012, at 14:19 , Richard Miles <> wrote:
> > I was not aware this Intel Xeon Phi, based on your slides I guess it will
> > be much more efficient for password cracking in comparison with the current
> > AMD GPU,

This will vary by hash type.  For bcrypt, yes.  For raw MD5, no.

> > however I doubt it will have a low cost such as a AMD GPU, right?

I can't answer this, other than by pointing at the link magnum already
found and posted (suggesting $400 per card for the first supercomputer
to use these).  We'll see what the retail pricing will be, if any.

> > Consequently it will not be affordable for most individuals and it will
> > decrease the number of implementations supporting it (example john the
> > ripper, hashcat, etc).

Maybe.  We'll see.

> > So, I guess it may not be very useful for most of
> > us, more or less the same that happens with PICO computing.

Pico's smallest FPGA board is $400 retail.  There are similarly priced
FPGA boards from other vendors as well, and cheaper ones too.  You do
need a bigger and much pricier board to get decent performance, though.

Price is just one of the factors.  There are other factors that prevent
FPGAs from being used more widely.  Consumer-oriented FPGA-based devices
can be competitive, as we see with FPGA-based Bitcoin miners.

> Maybe Phi won't be that expensive (beware your mail program might break the link in two):

Yeah, I first found this news story while working on my slides.

> Also, I assume we'll be able to use it with OpenCL.

A few months ago, when Xeon Phi was first announced under this name,
Intel supported OpenMP for it, but not OpenCL yet.  An Intel person
commented on a forum that OpenCL would likely be available later.  So I
doubt that there's OpenCL for these yet.

> As soon as there's drivers for it, we already have an implementation supporting it! Then again, we do not have code really optimised for it.

Most likely, we'll need to add intrinsics for the 512-bit vectors, and
build with OpenMP or run separate processes - right on the Xeon Phi.
It will run its own copy of Linux.  Even x86 real mode is there - in
each of the 50+ cores - to support bootup of existing OSes!  Apparently,
the die area cost for this was negligible (or at least acceptable).

The main selling point for Xeon Phi vs. GPUs appears to be that it's
more CPU'ish, including as it relates to porting of existing software
that does not use CUDA/OpenCL yet.  In this context, it makes a lot of
sense that Intel is supporting C/C++/Fortran with OpenMP and intrinsics
first, and OpenCL later.  BTW, the 512-bit intrinsics were specified for
Larrabee a few years ago; I did not check closely, but my expectation
is that MIC intrinsics are pretty much the same.  This should include
scatter/gather - Larrabee had it - hence the expectation for ability to
implement bcrypt efficiently.


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.