Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 18 Jul 2013 09:55:56 +0530
From: Sayantan Datta <>
Subject: Re: md5 hash comparisons


On Thu, Jul 18, 2013 at 6:00 AM, Michael Samuel <> wrote:

> I had quite a bit of success with this approach with some otp-md5
> stuff I did, first because branches on GPUs are(/were?) really
> expensive and second because gpu->host transfers are async so the gpu
> can be computing the next batch while the transfer occurs.

This could have been a good idea for small number of hashes. But as solar
mentioned we will be checking more than a million hash against the millions
generated. I think a single cpu core will bottleneck at some point.

What I intend to do is use local memory for first 2 bitmap checks each of
size 4KB so that they fit inside LDS . Then the last two bitmap check will
use bitmap stored in global memory each of size 256MB.

However I did a little math which shows 99% of hashes are going through the
first two test. Mathematically the whole bitmap idea seems to be
ineffective. But the plus point is that the math I did was based on the
assumption that all possible hashes md5 hashes are loaded. But clearly this
will never be the case.


Content of type "text/html" skipped

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.