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  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]
Date: Tue, 26 Aug 2014 02:37:23 +0400
From: Solar Designer <>
Subject: Re: [PHC] Re: Tradeoff cryptanalysis of password hashing schemes


I'm no expert in this, but:

On Mon, Aug 25, 2014 at 12:32:58PM -0400, Bill Cox wrote:
> A reasonable goal would be to have several such GDDR interfaces on
> your ASIC.  It is very hard to do!  However, it is doable, at least in
> theory.  If you had 512 pins all running at 6Gpbs, you'd have 384GiB/s
> bandwidth.  That would probably set a world record for chip bandwidth,
> but it's within a factor of 2X of what has been built before.

This is very close to what high-end GPUs and Xeon Phi currently (at
least claim to) achieve - on a 512-bit bus with GDDR5 memory, yes.
Current top GPUs are 288 GB/s to 320 GB/s.  Xeon Phi 5110P is 320 GB/s,
7120P is 352 GB/s.  Yes, these are GB/s, not GiB/s, so your 384 GiB/s
is still ~17% faster.  I think you were subtly wrong about it, though:
6 Gbps probably uses powers of 10, not of 2.  Only 9% faster than
Xeon Phi 7120P, then.  So it's clearly within reach, perhaps with some
overclocking of the lucky ones of those same chips.  I guess some
scrypt-based cryptocoin miners have already been overclocking their GPUs
like that... even though it's 20% higher than the GPUs' 320 GB/s at
stock clocks.  Whether such overclocking is a reasonable thing to do is
another matter, but this being the new world record is fairly unlikely.

What do you think of HMC?  A few months (or even a year?) ago, @jangray
tweeted pictures of prototype FPGA+HMC boards from a technology
exhibition he attended.  I don't recall this reliably now, but I guess
those boards used individual HMCs, so 160 GB/s only (half of what we
have with GPU boards), but perhaps this can be scaled to accessing
multiple HMCs from a single FPGA or ASIC?  How many?

"according to Micron, a single HMC can produce an incredible data
bandwidth of 160GBytes/sec"
"demonstrated interoperability between Micron's HMCs and Altera's
Stratix V FPGAs using a full 16-transceiver HMC link"

I guess a relevant question is: how many suitable transceivers do
current FPGAs have?  From the above, clearly it's at least 16, but if
it's 32+, then 2+ HMCs can be interfaced at full speed from one FPGA,
right?  And how many can an ASIC reasonably have?  Also, would the cost
and/or power consumption of the HMCs dominate anyway, in which case
lowering the number of ASICs wouldn't be all that important?

Four FPGAs and one 160 GB/s HMC on a board, so not that impressive for
our needs yet - but shows that boards with HMCs are already commercially
available.  And given demand, I guess boards with at least as many HMCs
as FPGAs on them can appear as well.


Powered by blists - more mailing lists