phc-discussions - Re: [PHC] Memory-hard proof of work with fast verification (CPU Hash)

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

Message-ID: <20150704160649.GA23270@openwall.com>
Date: Sat, 4 Jul 2015 19:06:49 +0300
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Memory-hard proof of work with fast verification (CPU Hash)

On Sat, Jul 04, 2015 at 03:56:57AM -0700, Bill Cox wrote:
> random ROM (maybe 512MiB), and a fairly small RAM (20KiB).
[...]
> Defeating GPU attacks is with the large ROM.  Scrypt achieves parity
> somewhere around 4MiB.  This would be much larger, maybe 512MiB.

We shouldn't compare RAM vs. ROM sizes like that.  RAM is per instance,
ROM is shared.  Having a 512 MB ROM will, on its own, have only small
impact on performance on modern GPU cards.  For example, on a card with
4 GB RAM, it leaves 3.5 GB for the instances' "RAMs", thereby reducing
concurrency to 7/8 of what would be possible without a ROM.  The
expected performance is thus somewhere in the range of 7/8 to 1 of
what it would have been without a ROM.

To defeat GPUs via a large ROM, you need to make it larger than the GPU
cards' RAM sizes, so some ROM accesses will have to go over PCIe
(transferring ROM blocks from other GPU cards' RAM or from host's RAM).

And yes, someone might want to introduce a cryptocoin based on yescrypt
with ROM enabled.  Maybe its ROM size should even grow over time, e.g.
starting at 28 GB (4 GB higher than what current top GPU cards use, and
4 GB lower than what fits in current desktop systems) and doubling every
few years or something... and eventually this might favor HPC clusters
with fast interconnect or something.

Alexander