Hardware purchase choice for NFS?

[VBCurtis]

I am nearing time to replace my desktop. An i7 setup is $100 more than an i5 of similar speed & overclocking options (both Haswell). Which would you choose for a machine that will run NFS factorizations for most of its life, and why?
Perhaps more specific questions are in order: For desktop CPUs, what increase in sieving throughput does hyperthreading produce? Is there a power-use difference for an i5 and i7 at the same speed? I am an experienced but conservative overclocker, if that matters.

How much slower are AMD chips than a Haswell for sieving (or linalg, if you know)?
-Curtis
Followup: Is DDR3-1600 enough memory bandwidth?

[xilman]

Originally Posted by VBCurtis

I am nearing time to replace my desktop. An i7 setup is $100 more than an i5 of similar speed & overclocking options (both Haswell). Which would you choose for a machine that will run NFS factorizations for most of its life, and why?
Perhaps more specific questions are in order: For desktop CPUs, what increase in sieving throughput does hyperthreading produce? Is there a power-use difference for an i5 and i7 at the same speed? I am an experienced but conservative overclocker, if that matters.

How much slower are AMD chips than a Haswell for sieving (or linalg, if you know)?
-Curtis
Followup: Is DDR3-1600 enough memory bandwidth?

Are you going to be concentrating on small factorizations or large ones? Large ones require much more memory for the LA. Actually, they require more memory for the sieving too but that's not usually a limitation these days.

My experience is that for sieving the more cores the better but hyperthreading is effectively useless. Productivity is effectively the product of the number of cores and the GHz of each core for a given size of cache. AMD performs pretty much the same as Intel by this metric; what's more important is the size of the L2 cache.

If you are going to be running GNFS get a half-way decent CUDA card for polynomial searching. It doesn't have to be anything special as even a second rate card will blow the cpu out of the water. I invariably search on an old laptop because its GT240M is markedly faster than any of my systems' cpus. It's hard to get a card that slow these days.

If you want high sieving performance it can be cost effective to pick up old systems dirt cheap, especially if you also need to heat your home where you live.

Sorry I can't give you specific advice about Haswell chips bu I've no experience with them. I've been running NFS for about 15 years though and have picked up quite a few rules of thumb which may be useful to you when you get up and running.

Paul

[fivemack]

Originally Posted by xilman

My experience is that for sieving the more cores the better but hyperthreading is effectively useless.

My experience with Haswell is that hyperthreading gains you about 30%; eight threads complete in six hours whilst four threads complete in four (http://www.mersenneforum.org/showpos...37&postcount=4). I would not buy a non-hyperthreaded Haswell.

I have a rather odd AMD machine, and find that one thread of i7/4770 gets me about the performance of 1.5 CPUs of K10/1900 - the one Haswell box has a quarter the performance at sieving of the whole 48-core AMD box.

Tom

[VBCurtis]

I have a quad-i7 laptop with a CUDA card, which is where I've done most of my NFS work. I found HT to produce 25-30% speed gains on that machine, but suspected that had to do with the relatively faster memory performance relative to CPU speed (1333[?] memory, 1700 CPU). I'm glad to hear the desktop realm offers the same percentage increase- that makes the i7 a clear choice so long as I'm doing ECM or NFS work.

From the NFS@Home forum, I gather that 30-bit projects can be solved in 4GB of memory, 31-bit in 6-8GB (if not, sieve more) for the linear algebra. Sounds like 8GB memory is sufficient for any task I could sieve on my own with 10 desktop cores?

Thank you for the opinions.

[debrouxl]

Sounds like 8GB memory is sufficient for any task I could sieve on my own with 10 desktop cores?

Yup.