Date: Thu, 2 Apr 2015 14:55:46 +0300
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Compute time hardness

Disclaimer: this is really not my area.  I would have preferred to let
someone else argue this.  But we don't appear to have anyone with just
the right expertise in here.  Bill's expertise is probably the closest
match among this community to what we'd need for this discussion.

On Thu, Apr 02, 2015 at 09:21:35AM +0200, Dmitry Khovratovich wrote:
> On Wed, Apr 1, 2015 at 11:21 PM, Solar Designer <solar@...nwall.com> wrote:
> >
> >> But anyway, as 32-bit multiplication is equivalent to a sum of 16 (average) addends, it can not be more than 4x slower given sufficient area (which we have).
> >
> > What are you averaging here?  I think we shouldn't.  I think it's 32.
> 
> 16 average non-zero addends. Very likely that no more than 24.

You mean 16 average non-zeroes (across the 32 partial products) in any
one bit position?  Yes, that's likely, but so what?  We don't know
which ones are zero vs. not when we design the circuit - this will vary
by its inputs.  So the delay is there all the time.

> So 4 or 5x extra depth assuming no other optimizations.

5x, plus significant overhead.  As to possible other optimizations,
that's interesting.  It'd be helpful to find any actual results on that.

> However, it is clear
> that the least significant bits of the final sum can be computed using
> only LSBs of intermediate sums, which makes the delay even smaller.

I specifically considered this aspect when designing pwxform (there was
even a thread in here on that very topic, which I started).  It tries to
ensure the MSBs determine the overall delay of multiple rounds of
pwxform.  The same can't be said of Blake2b's ADDs - there was no
deliberate attempt to have their MSBs determine the overall delay.

> > So I think it's 1 AND + 5 ADD.  And we shouldn't ignore that feeding
> > each input bit to 32 ANDs probably requires buffers.
> 
> We can assume that every bit has 32 copies, the area is still negligible.

You mean by computing 32 independent instances of the inputs to the
multiplier?  OK, but those instances' inputs come from a previous round
of the same primitive, also involving a multiplier, and thus also
needing 32 copies of their inputs each to avoid the original problem.
Even if this chain is broken between pwxform invocations (just not
between pwxform rounds), for 6 rounds we get over a billion of
instances.  OK, it doesn't really have to be 32 copies - perhaps we can
drive e.g. two or four AND gates from one output of a previous round
with little additional delay introduced, and perhaps we can introduce
some buffers before some of the multipliers (just not all) to break this
chain.  But this is still significant overhead.  Also, the larger the
circuit is, the larger the capacitances and the longer it takes for
gates' inputs to settle.

> > In that PhD thesis you referenced above, see page 136 (page 151 in PDF).
> > It shows the summation network corresponds to 49% of total delay in
> > their chosen multiplier.  In fact, it says: "The delay effects of the
> > other pieces of a multiplier are at least as important as the summation
> > network in determining the final performance."
> 
> That's for the particular (Booth-3) multiplier scheme, not for a naive
> one where you only have a binary tree of additions.

Fair enough.  The contribution of various components of a multiplier to
its overall delay may be different.

I guess the naive multiplier is slower overall, and I guess this is so
obvious to experts they don't even talk about it, but it'd be nice to
have this confirmed or disproved by an expert.

The naive multiplier appears faster if we assume that gate delays are
fixed and there are no other delays, and we merely count the circuit
depth - but in practice many other factors affect the overall delay as
well.  If the naive multiplier were faster, then wouldn't it be listed
as the fastest or at least used as a baseline in that PhD thesis?  The
fast multipliers given there are actually very large - I think they're
on par with or even larger than the naive one.

For example, their chosen multiplier design (not the absolute fastest,
but also not the largest and not the most power hungry), achieving 3.5 ns
typical (4.4 ns worst), "has about 60,000 bipolar transistors", and
has die area 10 mm^2, and consumes 4.5W (at the old 0.6 um process).
The fastest is 2.6 ns and is ~1.5x bigger and ~2x more power hungry.

BTW, I am impressed they could effectively match a 1 GHz CPU's multiply
latency at that old process.  Also, the use of primarily bipolar
transistors is unusual.

Can we estimate the number of transistors needed for a naive multiplier
(minimizing solely the circuit depth)?  60,000 / 1024 partial product
bits gives us 60 transistors per bit.  Perhaps half of them would be
spent on the first set of additions (since 1024 is close to
512+256+128+64 = 960 inputs to further additions).  If so, we have 30
transistors to spend on one full adder, which is more than we need -
which is 8 to 22, from what I could find.  Of course, we wouldn't use
ripple-carry adders, but 30 transistors per bit feels about right for a
carry look-ahead adder.

So the circuit size appears on par with the naive one (and their fastest
one is probably larger than the naive one).  Clearly, so much effort was
put into this not just for the sake of weirdness.

Alexander

Powered by blists - more mailing lists