CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW)

[James Heinrich]

Originally Posted by owftheevil

I also have results for a 570, 780, Titan, and Titan black if you want them.

Thanks. More benchmarks are required, I think. Between the three 3.5 cards the output devitates +/- 25% from my expected. The Titan Black is right on the mark, the 780 is 30% slower and the Titan is 20% faster than expected.

@LaurV (and anyone else who has one), if you can send me your 780/Titan* benchmarks (2.05) sometime that would be great.

[owftheevil]

Originally Posted by LaurV

Your Titan/Titan Black are about 1-2% faster than mine .
Were they done with P95 running? (mine yes, I try to go for "as close to real conditions as possible" when I build them)
Or you keep them cooler? (how? mine are air cooled and with hot Thai days they lose a lot of productivity when the aircond is off, for example, the TDP go down, and the iterations go from 3 to 4 ms, or even 5 etc.)
Just curious.

These were all done with mprime running. Mine are water cooled and go from 37--43 C depending on room temperature. That's probably the difference. The titan black settles in at 1019 Mhz and the regular titan at 923--947 Mhz.

[owftheevil]

Originally Posted by MacFactor

Please see this post:
http://www.mersenneforum.org/showthr...570#post396570

Two others that had this same problem solved it by building their own binaries locally. Its easy, but involves a large download (most of which is unnecessary, but unavoidable) if you don't have all of the cuda toolkit already.


Has anyone running Linux been able to get the binaries provided on Sourceforge working?

[LaurV]

Originally Posted by owftheevil

These were all done with mprime running. Mine are water cooled and go from 37--43 C depending on room temperature. That's probably the difference. The titan black settles in at 1019 Mhz and the regular titan at 923--947 Mhz.

Mine stabilizes at ~900 during the day, and ~970 during the night, with the TDP between say, 60 and 75%. This is the "worst" period of the year, because it is getting hot during the day, close to 30C, but we have around 16C during the night, sometimes higher. So, with windows open in the morning/night, the temperature in the rooms during the day stays under 25C and is not so hot to run the aircond. So, it is "not so hot, no so cold". When it is hot outside, as from end of March to June or so (before the raining season starts) we run the aircond in the house so it is better for the air cooled cards, somehow, during the day. Well, it is not really better, but the day/nigh difference is not so big.

Also, for the "water cooled" part, this period is bad because the sun is down and it shines under the roof of the cabinet, heating a corner of one of the radiators, therefore there are bigger temperature differences between day and night. For example, one water loop with one 2600k CPU and two 580's can reach as high as 60C during the day and as low as 28C during the night. In April, the hottest period in this area (45C during the day and 30C in the night) the temperature of the water plays around 45-50 during the day, in spite of the higher atmospheric temperature, because the sun is high in the sky and all the water cabinet is in the shadow.

I have to move my ass one of these days and mount that Titan water cooling blocks I was talking about last year... it's getting rusty in the drawers.. (I did not change it as long as the air cooler worked, in spite of the fact that I bought the water block after I got the first Titan, from Xyzzy, years ago. That one is still running, and on air, see my posts in my cooling thread here around).

[Mark Rose]

Originally Posted by LaurV

For example, one water loop with one 2600k CPU and two 580's can reach as high as 60C during the day and as low as 28C during the night.

What do you have in the way of radiators on that loop? I am curious.

[LaurV]

offtopic moved here

[Karl M Johnson]

Does anyone have a GTX 780 Ti and is willing to benchmark it with CUDALucas v2.05.1(CUDA 6.5) ?
I was shocked to learn that it might be faster than GTX Titan (while NV took away the option to enable 1/3 FP64 performance vs 1/8 default, according to one source).
The big question is: does GTX 780 Ti beat GTX Titan, even if the latter has 1/3 FP64 performance enabled?

For comparison, it takes CUDALucas 2.12 ms to perform 1 LL iteration(last two digits cut off) on M57885161, using GTX Titan.
Disabling double precision in NV CP slows CL to 3.62 ms / iteration.
This suggests that 780 Ti should be slower than GTX Titan, even if it has more shaders and a higher vRAM frequency.
Or not?

Anyways, if someone's up for the task, make sure to run threadbench and FFTbench, for best parameters:

CUDALucas2.05.1-CUDA6.5-Windows-x64 -cufftbench 1024 8192 3
CUDALucas2.05.1-CUDA6.5-Windows-x64 -threadbench 1024 8192 3 1

Over and out.

[axn]

Originally Posted by Karl M Johnson

For comparison, it takes CUDALucas 2.12 ms to perform 1 LL iteration(last two digits cut off) on M57885161, using GTX Titan.
Disabling double precision in NV CP slows CL to 3.62 ms / iteration.

Mersenne.ca gives 3.0ms/it for a 55m expo for Titan! That suggests that perhaps it is the Titan data that needs to updated at the site.

[Karl M Johnson]

Originally Posted by axn

That suggests that perhaps it is the Titan data that needs to updated at the site.

GTX Titan is 2 years old, CUDALucas got faster during that time.
I selected CUDA 6.5 binary since it had the best timings during the FFT bench (wondering if it's the same for all GPUs with shader model 3.5?)

[James Heinrich]

I've twiddled how the performance is shown for the (original) Titan based on how a recently submitted benchmark lines up with the other Titan benchmark I have. What I need now is more benchmarks from both 780 and Titan Black to figure that part out too. Benchmarks from Titan Z and Titan X would great too, but I'm not sure how many people have those available.

[Karl M Johnson]

Got some fresh info about GTX 780 Ti.

69M exp
1045 MHZ GPU clock, 3196 MHz mem clock, iteration time: 4.22 ms/iter
1176 MHz, 3570 MHz mem clock, iteration time: 3.7 ms/iter

With my GTX Titan, I get 2.76ms/iteration on the same exponent.

Promising data