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

[MooMoo2]

Originally Posted by nucleon

Has there been any work done on the competition?

On ATI's cards - any work done on them?

Probably not. It doesn't really matter anyway, since an average dual core CPU should be able to outperform either a Nvidia card or an ATI card:

http://www.mersenneforum.org/showpos...82&postcount=1

Q: So I can look for prime numbers on a GPU now?

Indeed you can.

Q: So how fast does it go?

It's a work in progress, but with a top-of-the-line card the current speed seems to be around what one core of a high-end PC can achieve.

[TheJudger]

Originally Posted by MooMoo2

Probably not. It doesn't really matter anyway, since an average dual core CPU should be able to outperform either a Nvidia card or an ATI card:

This might be a bit outdated. msft did some nice improvements!

http://www.mersenneforum.org/showpos...&postcount=206

A Core i7 980x (topend 6core) is around 7.5ms for a 2M FFT using all 6 cores.

Oliver

[nucleon]

GTX 480 with ver R code, 2M FFT 4.39ms.

As per TheJudger, the faq was written pre Fermi architecture. Top end GTX480 is streaks ahead in raw numbers, output per initial costs and output per ongoings.

The second fastest computer on the planet is using GPGPUs based on fermi:

www.top500.org

I'm sorry, but the time as come where we can no longer dismiss GPGPUs.

-- Craig

[The Carnivore]

Can someone tell me what the FFT length requirements are for CPUs vs GPUs? For example, if you wanted to check M24036583, you'd need a FFT length of 1280K if you were using a CPU. What FFT length would you need if you were using a GPU? 1280K? 2048K? 4096K?

I remember reading somewhere that starting a prime searching project without GPU support was like going to war without a piano. Some GPUs may match an overclocked quad or hex core if you were to compare FFT lengths, but they won't be of much use if they need FFT lengths that are much longer than CPUs.

BTW, my friend has a high end GPU, and he says it adds nearly 500 watts to his system at full load. Yuck.

[frmky]

Originally Posted by The Carnivore

For example, if you wanted to check M24036583, you'd need a FFT length of 1280K if you were using a CPU. What FFT length would you need if you were using a GPU? 1280K? 2048K? 4096K?

BTW, my friend has a high end GPU, and he says it adds nearly 500 watts to his system at full load. Yuck.

Round up to the nearest power of 2. 1280K > 1024K so you have to use 2048K FFT. But remember that a GTX 480 runs at under 4.5 ms/iteration using the 2048K FFT. Compare that with your CPU's speed at 1280K.

A PCIe card is limited to drawing 300W at full load by spec, but the cards are coming close to that limit.

[The Carnivore]

Originally Posted by frmky

Round up to the nearest power of 2. 1280K > 1024K so you have to use 2048K FFT. But remember that a GTX 480 runs at under 4.5 ms/iteration using the 2048K FFT. Compare that with your CPU's speed at 1280K.

My core i7 920 overclocked to 3.2GHz gets me about 20ms/iteration. But that's with only one core running. With all cores running one LL test each, the total output matches the GTX 480.

Surpassing the GTX 480 should be possible if you get an i7 980x or use more aggressive overclocks.

[Mini-Geek]

Originally Posted by The Carnivore

My core i7 920 overclocked to 3.2GHz gets me about 20ms/iteration. But that's with only one core running. With all cores running one LL test each, the total output matches the GTX 480.

But keep in mind that the GPU needs to go to the next highest power of 2. Going from 1280K to 2048K is approximately the worst-case scenario, since 1280K is just a step above 1024K, so the GPU has to take a huge performance hit (compared to a 1024K number) to run this number, while the CPU only had to take a small performance hit (compared to a 1024K number) . Of course, the best case would be where they both would use a power-of-2 base. So in a worst-case (for the GPU) scenario, you still need all cores of your i7 working together to match its output! In a best-case scenario, it's closer to twice as fast as your CPU.

[ET_]

Originally Posted by Mini-Geek

But keep in mind that the GPU needs to go to the next highest power of 2. Going from 1280K to 2048K is approximately the worst-case scenario, since 1280K is just a step above 1024K, so the GPU has to take a huge performance hit (compared to a 1024K number) to run this number, while the CPU only had to take a small performance hit (compared to a 1024K number) . Of course, the best case would be where they both would use a power-of-2 base. So in a worst-case (for the GPU) scenario, you still need all cores of your i7 working together to match its output! In a best-case scenario, it's closer to twice as fast as your CPU.

I'd add that with his computer AND the GPU, he could run 2 jobs in parallel...

Luigi

[sanzo]

Hello everyone,

I'm new member of this forum but I'm a GIMPS member from an year.

I have two graphics card
-an ATI RAdeon HD 5870
-and a Geforce 210

I found this discussion because I wont to test Lucas-Lehmer test on a GPU.

I want to know if there is a opencl program that can run on ATI's card or if there is only a cuda version (for my 210 card)?

Thanks all
sanzo

[henryzz]

Originally Posted by sanzo

Hello everyone,

I'm new member of this forum but I'm a GIMPS member from an year.

I have two graphics card
-an ATI RAdeon HD 5870
-and a Geforce 210

I found this discussion because I wont to test Lucas-Lehmer test on a GPU.

I want to know if there is a opencl program that can run on ATI's card or if there is only a cuda version (for my 210 card)?

Thanks all
sanzo

Only CUDA so far.

[sanzo]

Thanks henryzz,

Can I test this program on my gf 210??
I've win 7 64 bit, how I can try it? (exist a precompiled win version)...

sanzo