CUDA-EECM?

[ixfd64]

Has anyone ever used CUDA-EECM? Apparently, it's a CUDA-enabled version of GMP-ECM that was written by a Chinese researcher.

[xilman]

Originally Posted by ixfd64

Has anyone ever used CUDA-EECM? Apparently, it's a CUDA-enabled version of GMP-ECM that was written by a Chinese researcher.

I looked into it a while back. Unless it has now been extended, it's effectively useless as a general purpose factoring tool because it's limited to 300 bit arithmetic or so.

However, it could be rather useful as a co-factorer in a NFS and/or QS implementation.

Paul

[diep]

Originally Posted by ixfd64

Has anyone ever used CUDA-EECM? Apparently, it's a CUDA-enabled version of GMP-ECM that was written by a Chinese researcher.

I may hope for the guy he programs for AMD, as those gpu's are 5 times faster than nvidia's for stuff like that.

The new nvidia gpu that still has to release (fermi) will be real expensive and delivers single precision 1 Tflop.

In the shop you can already find a sapphire 5970 gpu from AMD (ATI) which at 850Mhz delivers 4.6 Tflop and in fact sapphire clocked it a tad higher. It's eating a whopping 300 watt nearly at full usage, but clocked nearly to a Ghz it's worth it. A bit under 1000 euro it is.

Thanks,
Vincent

[diep]

Originally Posted by diep

I may hope for the guy he programs for AMD, as those gpu's are 5 times faster than nvidia's for stuff like that.

The new nvidia gpu that still has to release (fermi) will be real expensive and delivers single precision 1 Tflop.

In the shop you can already find a sapphire 5970 gpu from AMD (ATI) which at 850Mhz delivers 4.6 Tflop and in fact sapphire clocked it a tad higher. It's eating a whopping 300 watt nearly at full usage, but clocked nearly to a Ghz it's worth it. A bit under 1000 euro it is.

Thanks,
Vincent

p.s. AMD will keep that lead of course (or better ATI), as it scales better that gpu. It has 3200 streamcores versus nvidia 448 for a chip that still has to release. Note the AMD has 2 gpu's on the board, but that doesn't matter to you, you care for the throughput of it and versus which price, isn't it?

[xilman]

Originally Posted by diep

I may hope for the guy he programs for AMD, as those gpu's are 5 times faster than nvidia's for stuff like that.

The new nvidia gpu that still has to release (fermi) will be real expensive and delivers single precision 1 Tflop.

In the shop you can already find a sapphire 5970 gpu from AMD (ATI) which at 850Mhz delivers 4.6 Tflop and in fact sapphire clocked it a tad higher. It's eating a whopping 300 watt nearly at full usage, but clocked nearly to a Ghz it's worth it. A bit under 1000 euro it is.

Thanks,
Vincent

Fermi has been out for some months now. Example cards include GT480, 470, 465 and 460 and their variants. The prices really aren't that high, especially the 460 variants.

Paul

[jasonp]

With graphics cards there's a tipping point where the libraries we need to use are no longer bleeding-edge, and suddenly 'everybody' will realize that they can now program the things without tears. That tipping point happened with Nvidia cards about 18 months ago (I was pretty late to the party) but I don't think it's happened yet with ATI cards.

[Raman]

Is that Elliptic Curve method by using Edwards Curves much more efficient for finding out factors than that traditional (Montgomery / not affine) Elliptic Curves? Is there any software for finding out factors / running up with that elliptic curve algorithm by using Edwards curves? Has anyone tried off with using that out?

[jasonp]

Dan Bernstein has some analysis that shows Edwards curves find factors slightly more often than Montgomery curves, and their arithmetic is somewhat faster. If I remember right, he states that ECM is overall about 25% more efficient on average.

Bernstein and the researchers that work with him are the only ones that have completed a workable ECM implementation that uses Edwards curves. His earlier work modified GMP-ECM to be about 7-10% faster in stage 1, but all the current work focuses on graphics cards.

[Andi_HB]

In this old thread are links to papers from Birkner & Co.
http://www.mersenneforum.org/showthread.php?t=11106

[diep]

Originally Posted by jasonp

With graphics cards there's a tipping point where the libraries we need to use are no longer bleeding-edge, and suddenly 'everybody' will realize that they can now program the things without tears. That tipping point happened with Nvidia cards about 18 months ago (I was pretty late to the party) but I don't think it's happened yet with ATI cards.

Jasonp, what i'm trying to explain to you is that the AMD cards are 5 times faster single precision.

If you look to effective IPC which similar software to factorizing gets out of the cards, then Nvidia sits around 25%-30% ipc and AMD at about 50% ipc.

So that makes AMD nearly 10 times faster than nvidia.

Vincent

[frmky]

Originally Posted by diep

So that makes AMD nearly 10 times faster than nvidia.

Not quite. The HD 5970 has 3.5x the SP FLOPS as the GTX 480. This figure includes IPC. However, because the nVidia programming tools are currently better than AMD's and the Fermi GPU's have a true cache hierarchy, actual code rarely runs more than 2x as fast on the HD 5970.