View Single Post
Old 2013-07-08, 19:31   #16
lorgix's Avatar
Sep 2010

11478 Posts

Originally Posted by VBCurtis View Post
You're quite welcome. Now that I know you don't have a GPU, I volunteer to find you at least decent polys for the next few months of your factoring. They may not hit msieve's GPU-target range every time, but doing your poly selects beats LLR-cuda as a good use of my GPU.

Repeated searches in the C155+ region give me some incentive to learn how to test-sieve, which leads to fiddling with the knobs as Jason puts it. All useful knowledge, and possibly heading for a beginner-guide mid-level GNFS entry.

GPU ideas: Go cheap, just get anything that supports the right flavor of CUDA. Even a low-end current card will be 50-100x faster for stage 1, while using fewer watts than a $100+ card.
Well, I actually have a CUDA-capable GPU. But it is weak, and has few cores. And it is hooked up to my main screen.
There are currently five p^q-1-GNFS-candidates in the range 154-156. I could pretest some above 160 as well. I also have my eyes on a HP-number in the low 160s. So there's no shortage of candidates, if you're up for it.
I really should be getting a new GPU some time soon (in a few moths, at least). More powerful ones should be very useful for ECM. Does poly select saturate any GPU? (can it use any number of cores?)
Maybe I'll opt for a cheaper one; a GT630 based on GK208 (meaning 28nm) was released fairly recently.
Should I look for the latest Compute Capability? (I don't really know the significance of that)
Originally Posted by firejuggler View Post
The GTX5xx are 2 generation behind. It might be hard to get "new" one. I guess a GTX 650 or 640 (if they are not only for laptop) might work for you.
I was under the impression that the Fermi architecture was still very competitive in GPGPU. I'm not sure though.
The GTX650 should be close to peak performance/$ (in the sense of MHz*cores/$), in the 600-series, and so should be a good choice I suppose.
The 6- & 700-series use an entirely different architecture than the 4- & 500, so I don't really know how to compare them.
Specifically I'd like to know what correction factor I need in order to compare MHz*cores between Fermi and Kepler. In terms of poly select performance (and ECM performance).
lorgix is offline   Reply With Quote