Optimum LL settings+hardware for i5/i7
 

Hi all,

I was somewhat inspired by George's post in building a new dedicated i7 rig to ask about the optimum setup/settings for a dedicated prime crunching rig.
It's purely hypothetical at this stage if I'm going to build such a rig, but I'm such a hardware junkie that I'll use any excuse to get my hands on a shiny new CPU, and LL crunching seems a perfect excuse ;)

Anyways, first question is:

Does hyperthreading help LL tests?

If so, what's the best number of LL's per core with a CPU with/without HT?

So it comes down to an i5 750 vs a i7 860, the latter one being significantly dearer here in Australia.

Lastly, would a mild overclock, from say 2.66/2.80 to 3-3.2GHz affect the integrity of the LL tests?

Cheers :smile:

[quote=hj47;189932]Does hyperthreading help LL tests?[/quote]
Yes. In my experience, LL-like stuff tends to get about a 10% or so speed boost with HT.

[quote]If so, what's the best number of LL's per core with a CPU with/without HT?[/quote]
It depends somewhat on the CPU, but across the board (and with i5/i7's, I think), it generally follows this pattern: without hyperthreading, do one LL per core; with hyperthreading, do one LL per real core, but multithread it across two hyperthreads. Note that for the latter, you need to make sure that Prime95 is actually splitting the tests over two hyperthreads of the same CPU, rather than over multiple actual cores; for this, you need to use the AffinityScramble option, which I've never personally used and can't attest to the subtleties of its behavior. :smile:

Note that another option with hyperthreading is to simply run one LL per hyperthread--i.e., that would be 4 tests for an i5 and 8 tests for an i7. AFAIK, this should be just as fast (or at least very close to just as fast) as running one LL per core with each test split over two hyperthreads. The advantage of this is that you don't have to worry about AffinityScramble, and it works just as well for other work that can't be split across hyperthreads like LL can (such as TF).

[quote]So it comes down to an i5 750 vs a i7 860, the latter one being significantly dearer here in Australia.

Lastly, would a mild overclock, from say 2.66/2.80 to 3-3.2GHz affect the integrity of the LL tests?[/quote]
I'm afraid I can't answer your first question, not being particularly familiar with those two CPUs and how their configuration relates to LL testing. However, as for overclocking, I can attest that an overclock to the degree you're suggesting is quite fine for LL tests; even though it may be a bit of an apples-to-oranges comparison, I have a Q6600 that I've overclocked from 2.4Ghz to 2.8Ghz that runs LL-type work quite stably. (Note that I have a non-OEM cooler on this machine; for almost all overclocks, you'll want something better than the OEM cooler that comes with the CPU.)

At any rate, overclock it as far as you dare, but make sure you stress test it with Prime95 as you go along. My advice is to do a longer stress test for a heavier overclock. For my relatively mild overclock, I actually skipped the Prime95 stress test since the temperatures were staying quite low and the work I was doing on it (LLR tests for the No Prime Left Behind project that are much smaller than GIMPS's tests, so somewhat like Prime95's Small FFT stress test) stood up fine to spotchecking of residuals.

Various users have reported hyperthreading results in no performance gain to a very modest gain. Matt's 10% is more gain than most report.

Hyperthreading usually makes the CPU run hotter which may reduce your ability to overclock. To get the best bang for the buck, I'd recommend the i5 750.

[quote=mdettweiler;189941]Note that another option with hyperthreading is to simply run one LL per hyperthread--i.e., that would be 4 tests for an i5 and 8 tests for an i7. [/quote]

So you mean run 1 LL per thread (both logical and physical?)- thus as you said giving you 8 tests?

So in theory with an i7 you should get x2 the output of an i5 since you have double to amount of threads?

[quote=Prime95;189942]Various users have reported hyperthreading results in no performance gain to a very modest gain. Matt's 10% is more gain than most report.[/quote]
Ah, I see. I haven't had a hyperthreaded CPU myself for a while, and the 10% figure was just something I'd heard stated before as an upper bound for HT speed boosts from a P4. The figure I heard quoted was something in the ballpark of 6-10%.

BTW, it's actually Max, not Matt, though it's a surprisingly common mistake and you're not the first to make it. :wink:

[quote=hj47;189945]So you mean run 1 LL per thread (both logical and physical?)- thus as you said giving you 8 tests?

So in theory with an i7 you should get x2 the output of an i5 since you have double to amount of threads?[/quote]
Yes, one LL per thread for 8 total tests.

As for an i7 giving 2x the output of an i5, I've never had any, so I couldn't tell you from experience, though I would imagine that would be the case.

[QUOTE=hj47;189945]So you mean run 1 LL per thread (both logical and physical?)- thus as you said giving you 8 tests?

So in theory with an i7 you should get x2 the output of an i5 since you have double to amount of threads?[/QUOTE]

In theory : practice should work as well as theory predicts it should.

In practice it doesn't.

Running 8 simultaneous LL tests on an i7 will [U]not[/U] give out 2x more work than running 4 tests. Far from that. Actually, as we are using hyperthreading, and the gain is in the order of 6-10% as stated, what happens is that each of the 8 LL tests will take significantly longer (nearly twice as long) than if we were running just 4. Although I never tested it, I think the overall gain in throughput would be roughly what we gain via hyperthreading for a single core: 6-10% at the most.
But, as I remember having read somewhere that HT works much better on i7s than on older P4 architectures, I admit I may be wrong. It would be interesting to have some real numbers.

[quote=lycorn;189956]Running 8 simultaneous LL tests on an i7 will [U]not[/U] give out 2x more work than running 4 tests. Far from that.[/quote]
I don't think anyone's saying so. Maybe you're referring to this post: [quote=hj47;189945] So in theory with an i7 you should get x2 the output of an i5 since you have double to amount of threads?[/quote]
I think that yes, in theory an i7 could do double the work of an i5, but there are things such as memory bandwidth to consider. (which is completely quad- vs dual-core, not related to HT)
EDIT: Sorry, I was thinking an i5 was a dual-core with HT, not a quad without HT. I was wrong. Move along, nothing to see here

[quote=Prime95;189942]Various users have reported hyperthreading results in no performance gain to a very modest gain. Matt's 10% is more gain than most report.

Hyperthreading usually makes the CPU run hotter which may reduce your ability to overclock. To get the best bang for the buck, I'd recommend the i5 750.[/quote]

You most probably geht the most bang if you enable one/two other cores from the AMD Phenom 2 X2 550BE or X3 710 & 720BE [URL]http://mersenneforum.org/showthread.php?t=12364[/URL]

[quote=Mini-Geek;189972]I don't think anyone's saying so. Maybe you're referring to this post:
I think that yes, in theory an i7 could do double the work of an i5, but there are things such as memory bandwidth to consider. (which is completely quad- vs dual-core, not related to HT)[/quote]

You won't get double the power with ht because you don't have more cpu cores. Hyperthreading only gives you performance if one thread is waiting for I/O or other waitstates (which most probably happens if you have several programs runing so the OS can schedule it in a better way)...

[quote=Mini-Geek;189972]EDIT: Sorry, I was thinking an i5 was a dual-core with HT, not a quad without HT. I was wrong. Move along, nothing to see here[/quote]
Hmm, that's what I thought too. Is an i5 actually a quad? (If so, that would make the dualcore an i3 rather than an i5 as I thought.)