phc-discussions - Potential cost improvement over scrypt

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [day] [month] [year] [list]

Message-ID: <CAOLP8p5fJmbhhe=Q07W42VdnnfdA3XWGq9kDGTz88Z_Yhb9Pag@mail.gmail.com>
Date: Thu, 26 Dec 2013 11:17:59 -0500
From: Bill Cox <waywardgeek@...il.com>
To: discussions@...sword-hashing.net
Subject: Potential cost improvement over scrypt

I just did some basic benchmarks and math to see how much improvement we
can get over scrypt in terms of making it more expensive for custom
hardware to hash passwords.  I used my very cost effective Core i7 linux
box I built for $500, and estimated how much improvement I could get in
custom hardware.  The short answer is that a 100X cost reduction seems
plausible, using GDDR5 memory and custom ASICs.

Because of the 12-ish ns latency in GDDR5, I estimate a custom scrypt
cracker would take about 20ns per guess.  Assuming 1GB of hashing memory, I
estimate a cost of around $32/password guesser.  From my benchmarks, it
seem that scrypt can't be sped up as much as I thought originally, since it
reads/writes 64 bytes at a time, reducing the cache miss penalty, and in
any case the 3-ish cycles per byte of RNG data generation dominates.

My take away is that filling memory with random data needs to be as fast as
possible if we want to improve on scrypt's cost per guess protection.  Can
we build a secure password hasher that uses a highly insecure but
super-fast RNG?  To max out memory bandwidth, we need to be generating
closer to 4 bytes/cycle on multiple threads rather than 3 cycles per byte
on one.

Content of type "text/html" skipped