Re: [AMBER] GTX680

From: Scott Le Grand <>
Date: Wed, 28 Mar 2012 09:33:39 -0700

After a weekend with GTX 680, what I can say is that this is a great gaming
GPU with amazing single-precision and texture performance, but it has the
same overall memory bandwidth as a GTX 580 with significantly
*less*double-precision performance.

The upshot is that I expect it to only deliver 85-90% of a GTX 580. And
that's partially because there's no increase in memory bandwidth and mostly
because of the regression in double-precision performance. And that's a
shame because single-precision *screams* on this chip.

There are compensated single-precision accumulation algorithms that could
be used here to ameliorate the performance hit. But this is a dangerous
precedent to follow IMO that leads to code fragmentation because
double-precision on all Fermi-class GPUs was faster, more precise, and
simpler than such algorithms (which themselves where faster on GTX 2xx (see
Tetsu Narumi's work, sigh). This is a nightmare to validate: GTX 680
simply shouldn't have regressed on double-precision performance.
Hopefully, the next chip won't. That said, this thing overclocks like
crazy and the modder crowd has already doubled the base clock. So perhaps
not all hope is lost...


On Fri, Mar 23, 2012 at 11:57 AM, Scott Le Grand <>wrote:

> I am very optimistic about GTC680 performance...
> That said, anyone who hacks the configure script to make the current code
> run will be severely (an unnecessarily) disappointed. Every AMBER kernel
> has been meticulously shoehorned into GTX2xx and GTX5xx GPUs. GTX680 is a
> radical redesign of Fermi (please don't listen to the dunderheads on review
> sites blathering about matters that's beyond them about such things,
> seriously). That radical redesign has created a much more efficient GPU
> (I'm expecting the perf/watt on AMBER to hit transwarp as opposed to merely
> warp drive in the near future) but it's been at the expense of 33% higher
> operational latency.
> 33% higher operational latency is fine - except that the shared memory on
> GTX680 is exactly the same as GTX580 and that's leading to a ~30%
> performance deficit if one just runs the existing code. However, there are
> 2x as many machine registers on GTX680 than on GTX580. Or TLDR: I need to
> rewrite every single kernel for GTX680 from the ground-up to hit attainable
> performance.
> So give me a few weeks, mmkay?
> Scott
> On Fri, Mar 23, 2012 at 11:49 AM, Ross Walker <>wrote:
>> Hi Filip,
>> > Hi all,
>> > I was wondering
>> > what we can expect from GTX680 and in general from the new Kepler line.
>> > I know
>> > that GTX680 is very limited DP, but should be good in SP mode. Would we
>> > expect
>> > some speed boost compared to GTX580 and also will it work along Amber
>> > 11/12?
>> Amber 11 will NOT support the GTX680 cards (unless you hack the configure
>> script to compile it in what is effectively an emulation mode). It will be
>> too much work to make patch against that. AMBER 12 will support them but it
>> is going to take around 6 weeks to 2 months to get the optimization done
>> and a patch released so it won't support the cards at release but it will
>> as soon as we have the patch ready. I can't really give you any performance
>> expectations right now, only got my first prototype board yesterday. ;-)
>> Right now if you compile AMBER 12 with PTX support so that it will at
>> least run on the GTX680 the performance sucks. It is about 70% of a GTX580.
>> NVIDIA changed the hardware too much (massively increasing the threads but
>> also the thread latency) so it will need some work to optimize it which is
>> why I have chosen not to support the cards in AMBER 12 until we have that
>> optimization done. Once it is done I expect considerable improvement over
>> GTX580 speeds but can't give you anything concrete right now.
>> > P.S. Indeed most
>> > of us will probably wait for GK110, but CUDA capability of GTX680 is
>> > very limited now.
>> This is probably a good idea, at least you should wait until we have had
>> a chance to get our hands dirty with the GK104 chip. So I'd urge you to
>> wait at least until we have the patch ready for AMBER 12.
>> All the best
>> Ross
>> /\
>> \/
>> |\oss Walker
>> ---------------------------------------------------------
>> | Assistant Research Professor |
>> | San Diego Supercomputer Center |
>> | Adjunct Assistant Professor |
>> | Dept. of Chemistry and Biochemistry |
>> | University of California San Diego |
>> | NVIDIA Fellow |
>> | | |
>> | Tel: +1 858 822 0854 | EMail:- |
>> ---------------------------------------------------------
>> Note: Electronic Mail is not secure, has no guarantee of delivery, may
>> not be read every day, and should not be used for urgent or sensitive
>> issues.
>> _______________________________________________
>> AMBER mailing list
AMBER mailing list
Received on Wed Mar 28 2012 - 10:00:02 PDT
Custom Search